<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>Tech Adventures in Business Central | AL Development &amp; AI Solutions</title>
    <description>Exploring Business Central, AL coding, and AI-driven ERP solutions for developers and businesses.</description>
    <link>https://tine.staric.net/</link>
    <atom:link href="https://tine.staric.net/feed.xml" rel="self" type="application/rss+xml"/>
    <pubDate>Fri, 06 Nov 2026 11:05:23 +0100</pubDate>
    <lastBuildDate>Fri, 06 Nov 2026 11:05:23 +0100</lastBuildDate>
    <generator>Jekyll v4.3.2</generator>
    
      <item>
        <title>&quot;Dev X used 200k tokens with Opus&quot; is not telemetry</title>
        <description>&lt;p&gt;&lt;em&gt;“Dev X used 200k tokens with Opus.”&lt;/em&gt; That’s the kind of number you get out of the box, and it’s fine for a bill. It’s useless for deciding anything.&lt;/p&gt;

&lt;p&gt;What I actually wanted to know, after rolling &lt;strong&gt;&lt;em&gt;Claude Code&lt;/em&gt;&lt;/strong&gt; and &lt;strong&gt;&lt;em&gt;GitHub Copilot&lt;/em&gt;&lt;/strong&gt; out to a team with a shared set of plugins, skills and agents: which of those get used, and by how many people? How many people have a plugin installed but never touch it? What does a repository cost us? Is the thing we distributed used at all, or did we just make ourselves feel better?&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Claude Code&lt;/em&gt;’s built-in analytics gets you part of the way. It tells you that 10 of your users used plugins, but not which plugin:&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/harnesstelemetry/claude-analytics.png&quot; alt=&quot;Claude&apos;s built-in analytics: tool usage, and feature adoption as a percentage of active users&quot; /&gt;&lt;/p&gt;

&lt;p&gt;My first attempt at the rest was a hook that posts events to &lt;em&gt;Application Insights&lt;/em&gt; myself. It worked. Then I realized I’d own it forever: the schema, the auth, every harness update that renames an event.&lt;/p&gt;

&lt;p&gt;Then I found out both tools already emit &lt;em&gt;OpenTelemetry&lt;/em&gt;. A setup that works out of the box beats one I maintain, so I threw the hook away.&lt;/p&gt;

&lt;p&gt;This post has been a long time coming. I kept wrangling with what &lt;em&gt;Claude Code&lt;/em&gt;, the &lt;em&gt;Copilot CLI&lt;/em&gt; and &lt;em&gt;VS Code&lt;/em&gt; each emit, and &lt;em&gt;VS Code&lt;/em&gt; was painfully different from the other two. With the switch to the &lt;a href=&quot;/blog/2026/copilot-harness-default/&quot;&gt;new Copilot harness&lt;/a&gt; that’s finally over: VS Code and the CLI emit the same signals, so one setup covers both.&lt;/p&gt;

&lt;p&gt;Below is what the setup looks like, what to collect and what to leave out, what you get to look at, and how to roll it out to every machine. The collector configs, managed settings, queries and dashboards for Claude Code, Copilot and both together are in a &lt;strong&gt;&lt;a href=&quot;https://github.com/tinestaric/harness-telemetry&quot;&gt;sample repo&lt;/a&gt;&lt;/strong&gt;.&lt;/p&gt;

&lt;hr /&gt;

&lt;h3 id=&quot;otel-in-two-minutes&quot;&gt;OTEL in two minutes&lt;/h3&gt;

&lt;p&gt;&lt;em&gt;OpenTelemetry&lt;/em&gt; has three signal types: &lt;strong&gt;metrics&lt;/strong&gt; (numbers over time), &lt;strong&gt;events&lt;/strong&gt;, which it calls logs, and &lt;strong&gt;traces&lt;/strong&gt; (one span per operation, nested). That’s all you need. &lt;em&gt;Claude Code&lt;/em&gt; sends metrics and events. &lt;em&gt;Copilot&lt;/em&gt; sends traces and metrics.&lt;/p&gt;

&lt;p&gt;You could point both straight at &lt;em&gt;Application Insights&lt;/em&gt;, but it wants an &lt;em&gt;Entra&lt;/em&gt; token per developer, not a header you can put in a settings file. So there’s a small collector in the middle:&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/harnesstelemetry/pipeline.svg&quot; alt=&quot;Claude Code and Copilot send OTLP with a bearer token to your collector, which filters, hashes and exports to Application Insights, read by a Workbook or Grafana&quot; /&gt;&lt;/p&gt;

&lt;p&gt;The collector accepts a bearer ingest token, then filters, transforms and exports with the connection string. To be honest about that token: every client that receives the managed settings can see it. It only stops outsiders. I checked against the hosted collector: no token gets a 401, a wrong token gets a 401, the right one gets a 200.&lt;/p&gt;

&lt;p&gt;Hosting it is small. One &lt;em&gt;Container App&lt;/em&gt;, one container, one YAML file mounted from a secret, pinned to one replica. The ingest token, the salt and the connection string are secrets. One gotcha: changing only a secret doesn’t create a new revision, so restart it. The sample repo has the details, I won’t write a walkthrough here.&lt;/p&gt;

&lt;p&gt;Here’s a trimmed version of the config, with one receiver, one filter, one transform and one exporter:&lt;/p&gt;

&lt;div class=&quot;language-yaml highlighter-rouge&quot;&gt;&lt;div class=&quot;highlight&quot;&gt;&lt;pre class=&quot;highlight&quot;&gt;&lt;code&gt;&lt;span class=&quot;na&quot;&gt;receivers&lt;/span&gt;&lt;span class=&quot;pi&quot;&gt;:&lt;/span&gt;
  &lt;span class=&quot;na&quot;&gt;otlp&lt;/span&gt;&lt;span class=&quot;pi&quot;&gt;:&lt;/span&gt;
    &lt;span class=&quot;na&quot;&gt;protocols&lt;/span&gt;&lt;span class=&quot;pi&quot;&gt;:&lt;/span&gt;
      &lt;span class=&quot;na&quot;&gt;http&lt;/span&gt;&lt;span class=&quot;pi&quot;&gt;:&lt;/span&gt;
        &lt;span class=&quot;na&quot;&gt;endpoint&lt;/span&gt;&lt;span class=&quot;pi&quot;&gt;:&lt;/span&gt; &lt;span class=&quot;s&quot;&gt;0.0.0.0:4318&lt;/span&gt;
        &lt;span class=&quot;na&quot;&gt;auth&lt;/span&gt;&lt;span class=&quot;pi&quot;&gt;:&lt;/span&gt;
          &lt;span class=&quot;na&quot;&gt;authenticator&lt;/span&gt;&lt;span class=&quot;pi&quot;&gt;:&lt;/span&gt; &lt;span class=&quot;s&quot;&gt;bearertokenauth/ingest&lt;/span&gt;

&lt;span class=&quot;na&quot;&gt;processors&lt;/span&gt;&lt;span class=&quot;pi&quot;&gt;:&lt;/span&gt;
  &lt;span class=&quot;na&quot;&gt;filter/allowlist&lt;/span&gt;&lt;span class=&quot;pi&quot;&gt;:&lt;/span&gt;
    &lt;span class=&quot;na&quot;&gt;logs&lt;/span&gt;&lt;span class=&quot;pi&quot;&gt;:&lt;/span&gt;
      &lt;span class=&quot;na&quot;&gt;log_record&lt;/span&gt;&lt;span class=&quot;pi&quot;&gt;:&lt;/span&gt;
        &lt;span class=&quot;pi&quot;&gt;-&lt;/span&gt; &lt;span class=&quot;pi&quot;&gt;&amp;gt;-&lt;/span&gt;
          &lt;span class=&quot;s&quot;&gt;not (attributes[&quot;event.name&quot;] == &quot;skill_activated&quot;&lt;/span&gt;
          &lt;span class=&quot;s&quot;&gt;or attributes[&quot;event.name&quot;] == &quot;plugin_loaded&quot;&lt;/span&gt;
          &lt;span class=&quot;s&quot;&gt;or attributes[&quot;event.name&quot;] == &quot;subagent_completed&quot;)&lt;/span&gt;

  &lt;span class=&quot;na&quot;&gt;transform/pseudonymize&lt;/span&gt;&lt;span class=&quot;pi&quot;&gt;:&lt;/span&gt;
    &lt;span class=&quot;na&quot;&gt;log_statements&lt;/span&gt;&lt;span class=&quot;pi&quot;&gt;:&lt;/span&gt;
      &lt;span class=&quot;pi&quot;&gt;-&lt;/span&gt; &lt;span class=&quot;na&quot;&gt;context&lt;/span&gt;&lt;span class=&quot;pi&quot;&gt;:&lt;/span&gt; &lt;span class=&quot;s&quot;&gt;log&lt;/span&gt;
        &lt;span class=&quot;na&quot;&gt;statements&lt;/span&gt;&lt;span class=&quot;pi&quot;&gt;:&lt;/span&gt;
          &lt;span class=&quot;pi&quot;&gt;-&lt;/span&gt; &lt;span class=&quot;pi&quot;&gt;&amp;gt;-&lt;/span&gt;
            &lt;span class=&quot;s&quot;&gt;set(log.attributes[&quot;user.pseudo_id&quot;],&lt;/span&gt;
            &lt;span class=&quot;s&quot;&gt;Substring(SHA256(Concat([&quot;${env:PSEUDONYM_SALT}&quot;,&lt;/span&gt;
            &lt;span class=&quot;s&quot;&gt;log.attributes[&quot;user.email&quot;]], &quot;:&quot;)), 0, 16))&lt;/span&gt;
            &lt;span class=&quot;s&quot;&gt;where log.attributes[&quot;user.email&quot;] != nil&lt;/span&gt;
          &lt;span class=&quot;pi&quot;&gt;-&lt;/span&gt; &lt;span class=&quot;s&quot;&gt;delete_key(log.attributes, &quot;user.email&quot;)&lt;/span&gt;

&lt;span class=&quot;na&quot;&gt;exporters&lt;/span&gt;&lt;span class=&quot;pi&quot;&gt;:&lt;/span&gt;
  &lt;span class=&quot;na&quot;&gt;azure_monitor&lt;/span&gt;&lt;span class=&quot;pi&quot;&gt;:&lt;/span&gt;
    &lt;span class=&quot;na&quot;&gt;connection_string&lt;/span&gt;&lt;span class=&quot;pi&quot;&gt;:&lt;/span&gt; &lt;span class=&quot;s&quot;&gt;${env:APPLICATIONINSIGHTS_CONNECTION_STRING}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;&lt;/div&gt;&lt;/div&gt;

&lt;p&gt;Why &lt;em&gt;Application Insights&lt;/em&gt;? It was already there, it speaks KQL, and workspace access control comes with it.&lt;/p&gt;

&lt;hr /&gt;

&lt;h3 id=&quot;start-from-the-questions-not-the-signals&quot;&gt;Start from the questions, not the signals&lt;/h3&gt;

&lt;p&gt;The harnesses emit a long list of things: &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;skill_activated&lt;/code&gt;, &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;subagent_completed&lt;/code&gt;, &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;plugin_loaded&lt;/code&gt;, &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;mcp_server_connection&lt;/code&gt;, &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;tool_result&lt;/code&gt;, &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;cost&lt;/code&gt;. You could spend a week deciding which ones matter. The list doesn’t matter, the questions do. Mine:&lt;/p&gt;

&lt;ul&gt;
  &lt;li&gt;&lt;strong&gt;What adds value and what doesn’t?&lt;/strong&gt; Which skills, agents and plugins are used, and by how many people. Which are loaded but never used.&lt;/li&gt;
  &lt;li&gt;&lt;strong&gt;Where should we put effort?&lt;/strong&gt; How many people have a plugin but don’t use it. A plugin everyone loads and nobody touches needs a better description or a demo.&lt;/li&gt;
  &lt;li&gt;&lt;strong&gt;What does it cost?&lt;/strong&gt; Per repository, per model, and which sessions ran away.&lt;/li&gt;
  &lt;li&gt;&lt;strong&gt;Is the rollout healthy?&lt;/strong&gt; MCP servers failing to connect, client version drift.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The “8 of 10 people use the plugin” question is one KQL query: for each plugin, how many users and sessions loaded it, and how many skill uses came out of it.&lt;/p&gt;

&lt;pre&gt;&lt;code class=&quot;language-kusto&quot;&gt;let loaded = traces
| where timestamp &amp;gt; ago(30d) and message == &quot;claude_code.plugin_loaded&quot;
| summarize loaded_users = dcount(tostring(customDimensions[&quot;user.pseudo_id&quot;])),
            loaded_sessions = dcount(tostring(customDimensions[&quot;session.id&quot;]))
  by plugin = tostring(customDimensions[&quot;plugin.name&quot;]);
let used = traces
| where timestamp &amp;gt; ago(30d) and message == &quot;claude_code.skill_activated&quot;
    and tostring(customDimensions[&quot;skill.source&quot;]) == &quot;plugin&quot;
| summarize skill_uses = count()
  by plugin = tostring(customDimensions[&quot;plugin.name&quot;]);
loaded
| join kind=leftouter used on plugin
| project plugin, loaded_users, loaded_sessions, skill_uses = coalesce(skill_uses, 0)
&lt;/code&gt;&lt;/pre&gt;

&lt;p&gt;A plugin at the top with lots of sessions and zero skill uses is one people have but don’t use.&lt;/p&gt;

&lt;p&gt;One caveat on cost: &lt;em&gt;Claude Code&lt;/em&gt;’s cost figure is a client-side estimate in dollars, not a bill. &lt;em&gt;Copilot&lt;/em&gt; reports AI credits on its spans instead, and on the CLI runs I compared they matched what the CLI printed. I haven’t compared either against actual billing.&lt;/p&gt;

&lt;hr /&gt;

&lt;h3 id=&quot;collect-the-minimum-and-delete-the-rest-before-its-stored&quot;&gt;Collect the minimum, and delete the rest before it’s stored&lt;/h3&gt;

&lt;p&gt;The harnesses can emit a lot more than counts: prompts, responses, tool inputs, the user’s email. &lt;em&gt;Claude Code&lt;/em&gt; will let you track a user down to every single prompt they type. Try doing that in the EU and GDPR will likely strangle you. And I didn’t want it anyway. I wanted counts and names.&lt;/p&gt;

&lt;p&gt;One thing to know about the &lt;em&gt;Claude Code&lt;/em&gt; side: to see plugin, skill and agent names you have to turn on &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;OTEL_LOG_TOOL_DETAILS&lt;/code&gt;, and that also sends tool inputs: MCP queries, bash commands. Without it, every non-official plugin, skill and agent collapses to &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;third-party&lt;/code&gt; or &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;custom&lt;/code&gt;, which makes the dashboard pointless. So the collector has to be yours, over HTTPS, and it has to delete the content before anything is stored. I tested it by putting marker strings in an MCP query and a bash command. They showed up nowhere in the output.&lt;/p&gt;

&lt;p&gt;Two rules:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Allowlist.&lt;/strong&gt; The collector keeps only the events and metrics the questions need. Prompts and responses are off at the client. Lines of code, commits, PRs, tool approvals and API errors are dropped. Lines of code and the like are pointless KPIs, and I don’t want anyone mistaking the dashboard for a productivity score.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Hash the identity you can read, restrict the identity you can’t.&lt;/strong&gt; &lt;em&gt;Claude Code&lt;/em&gt; sends the email, so the collector replaces it with a salted hash. The salt lives only on the collector, never in managed settings. &lt;em&gt;Copilot&lt;/em&gt;’s &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;enduser.pseudo.id&lt;/code&gt; is already opaque, so I pass it through and let workspace access protect it. Same person, same id, so the counts still work without names.&lt;/p&gt;

&lt;p&gt;Pseudonymous is still personal data. Restrict access to the workspace, and say out loud what it’s for. I’d rather someone read my browser history than my prompt history.&lt;/p&gt;

&lt;hr /&gt;

&lt;h3 id=&quot;what-you-actually-look-at&quot;&gt;What you actually look at&lt;/h3&gt;

&lt;p&gt;You have options. &lt;em&gt;Azure Workbooks&lt;/em&gt; need nothing extra to host. &lt;em&gt;Azure Managed Grafana&lt;/em&gt; is the tool you can take elsewhere. &lt;em&gt;Power BI&lt;/em&gt; suits scheduled management reports. I tried Grafana first, and it looks nice:&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/harnesstelemetry/dashboard-claude-code-top.jpg&quot; alt=&quot;Grafana: 15 active users, plugins loaded vs used, skill usage, agents and MCP server connections&quot; /&gt;&lt;/p&gt;

&lt;p&gt;Then I went with &lt;em&gt;Workbooks&lt;/em&gt;. They run the same KQL, use the same Azure access control, and there’s no extra service to run. The sample repo has both versions.&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/harnesstelemetry/workbook-plugins.png&quot; alt=&quot;The Azure Workbook version of the plugins panel&quot; /&gt;&lt;/p&gt;

&lt;p&gt;Community dashboards exist for both tools, but they expose far too much. Once there are that many data points you stop looking at any of them, which is why I wrote my own queries. The first version of mine had 53 panels and had the same problem. The core dashboard now has nine: active users, plugins loaded vs used, skills, subagents, MCP connections, MCP tool usage, tool usage by name, cost by repository, and a check that no content got stored, which must read 0.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Department comes from the repository.&lt;/strong&gt; The email is hashed, so I map repository to department with a small lookup table.&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/harnesstelemetry/dashboard-cost-by-repo.jpg&quot; alt=&quot;Cost by repository with the department mapped in&quot; /&gt;&lt;/p&gt;

&lt;p&gt;(The repositories and departments in these screenshots are example data.)&lt;/p&gt;

&lt;p&gt;One thing the dashboard can’t tell you: which skills are &lt;em&gt;never&lt;/em&gt; used. Telemetry only reports what happened. To see the gaps you need a catalog of what each plugin ships, to compare against.&lt;/p&gt;

&lt;hr /&gt;

&lt;h3 id=&quot;getting-it-onto-every-machine&quot;&gt;Getting it onto every machine&lt;/h3&gt;

&lt;p&gt;Managed settings are the way, for a simple reason: they win over local settings, so developers can’t switch telemetry off, and you don’t need device management.&lt;/p&gt;

&lt;p&gt;For &lt;em&gt;Claude Code&lt;/em&gt; that’s the &lt;strong&gt;Managed settings&lt;/strong&gt; entry in the admin console, where I set it up.&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/harnesstelemetry/admin-console-managed-settings.png&quot; alt=&quot;The Claude Code section of the admin console, with the Managed settings entry at the top&quot; /&gt;&lt;/p&gt;

&lt;p&gt;It’s mostly an &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;env&lt;/code&gt; block, token redacted here, plus the known marketplace (more on that below):&lt;/p&gt;

&lt;div class=&quot;language-json highlighter-rouge&quot;&gt;&lt;div class=&quot;highlight&quot;&gt;&lt;pre class=&quot;highlight&quot;&gt;&lt;code&gt;&lt;span class=&quot;p&quot;&gt;{&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
  &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;extraKnownMarketplaces&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;p&quot;&gt;{&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
    &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;sample-marketplace&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;p&quot;&gt;{&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
      &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;source&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;p&quot;&gt;{&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
        &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;source&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;s2&quot;&gt;&quot;git&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
        &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;url&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;s2&quot;&gt;&quot;https://github.com/tinestaric/plugin-marketplace-sample.git&quot;&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
      &lt;/span&gt;&lt;span class=&quot;p&quot;&gt;},&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
      &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;autoUpdate&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;kc&quot;&gt;true&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
    &lt;/span&gt;&lt;span class=&quot;p&quot;&gt;}&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
  &lt;/span&gt;&lt;span class=&quot;p&quot;&gt;},&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
  &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;env&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;p&quot;&gt;{&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
    &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;CLAUDE_CODE_ENABLE_TELEMETRY&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;s2&quot;&gt;&quot;1&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
    &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;OTEL_METRICS_EXPORTER&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;s2&quot;&gt;&quot;otlp&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
    &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;OTEL_LOGS_EXPORTER&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;s2&quot;&gt;&quot;otlp&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
    &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;OTEL_EXPORTER_OTLP_PROTOCOL&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;s2&quot;&gt;&quot;http/protobuf&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
    &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;OTEL_EXPORTER_OTLP_ENDPOINT&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;s2&quot;&gt;&quot;https://&amp;lt;your-collector&amp;gt;&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
    &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;OTEL_EXPORTER_OTLP_HEADERS&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;s2&quot;&gt;&quot;Authorization=Bearer &amp;lt;ingest-token&amp;gt;&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
    &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;OTEL_EXPORTER_OTLP_METRICS_TEMPORALITY_PREFERENCE&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;s2&quot;&gt;&quot;delta&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
    &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;OTEL_METRICS_INCLUDE_REPOSITORY&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;s2&quot;&gt;&quot;true&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
    &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;OTEL_LOG_USER_PROMPTS&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;s2&quot;&gt;&quot;0&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
    &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;OTEL_LOG_ASSISTANT_RESPONSES&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;s2&quot;&gt;&quot;0&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
    &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;OTEL_LOG_TOOL_DETAILS&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;s2&quot;&gt;&quot;1&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
    &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;OTEL_LOG_TOOL_CONTENT&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;s2&quot;&gt;&quot;0&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
    &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;OTEL_LOG_RAW_API_BODIES&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;s2&quot;&gt;&quot;0&quot;&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
  &lt;/span&gt;&lt;span class=&quot;p&quot;&gt;}&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;}&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/div&gt;&lt;/div&gt;

&lt;p&gt;The only other managed setting I’d use is the known marketplace (&lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;extraKnownMarketplaces&lt;/code&gt;), so the shared plugin set is there without anyone adding it by hand. In the sample it points at the repo from &lt;a href=&quot;/blog/2026/agent-plugins/&quot;&gt;the plugins post&lt;/a&gt;, swap in your own. Permissions, MCP and plugin allow/deny lists, forced hooks, minimum version: all available, and I’d skip them for this.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Copilot&lt;/em&gt; has the same idea under a different name: &lt;strong&gt;enterprise managed settings&lt;/strong&gt;, with a &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;telemetry&lt;/code&gt; block that covers VS Code and the agent host behind the CLI. A managed value beats environment variables and user settings. I’ve only set up the &lt;em&gt;Claude Code&lt;/em&gt; side so far, so everything below is from GitHub’s docs.&lt;/p&gt;

&lt;div class=&quot;language-json highlighter-rouge&quot;&gt;&lt;div class=&quot;highlight&quot;&gt;&lt;pre class=&quot;highlight&quot;&gt;&lt;code&gt;&lt;span class=&quot;p&quot;&gt;{&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
  &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;extraKnownMarketplaces&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;p&quot;&gt;{&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
    &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;sample-marketplace&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;p&quot;&gt;{&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
      &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;source&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;p&quot;&gt;{&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
        &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;source&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;s2&quot;&gt;&quot;github&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
        &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;repo&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;s2&quot;&gt;&quot;tinestaric/plugin-marketplace-sample&quot;&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
      &lt;/span&gt;&lt;span class=&quot;p&quot;&gt;},&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
      &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;autoUpdate&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;kc&quot;&gt;true&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
    &lt;/span&gt;&lt;span class=&quot;p&quot;&gt;}&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
  &lt;/span&gt;&lt;span class=&quot;p&quot;&gt;},&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
  &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;telemetry&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;p&quot;&gt;{&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
    &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;enabled&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;kc&quot;&gt;true&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
    &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;endpoint&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;s2&quot;&gt;&quot;https://&amp;lt;your-collector&amp;gt;&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
    &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;protocol&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;s2&quot;&gt;&quot;http/protobuf&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
    &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;captureContent&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;kc&quot;&gt;false&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
    &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;lockCaptureContent&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;kc&quot;&gt;true&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
    &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;serviceName&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;s2&quot;&gt;&quot;github-copilot&quot;&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
  &lt;/span&gt;&lt;span class=&quot;p&quot;&gt;}&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;}&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/div&gt;&lt;/div&gt;

&lt;p&gt;The known marketplace works here too, with the same key. It also accepts a &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;github&lt;/code&gt; source (&lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;OWNER/REPO&lt;/code&gt;), which is shorter than the git URL.&lt;/p&gt;

&lt;p&gt;To set it up, you pick an organization in your enterprise to host the config (in the &lt;strong&gt;AI Controls&lt;/strong&gt; tab of the enterprise settings, or through the API), create the file &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;copilot/managed-settings.json&lt;/code&gt; in that organization’s &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;.github-private&lt;/code&gt; repository, and commit it to the default branch. The Agents page under AI controls shows whether it’s active. Clients fetch it when the user signs in and refresh it hourly. The same settings can also be pushed through MDM. We didn’t take that route, because it means a fight with the IT department.&lt;/p&gt;

&lt;p&gt;There are two caveats:&lt;/p&gt;

&lt;ul&gt;
  &lt;li&gt;&lt;strong&gt;It follows the account.&lt;/strong&gt; The server-managed route applies to users with a &lt;em&gt;Copilot Business&lt;/em&gt; or &lt;em&gt;Copilot Enterprise&lt;/em&gt; seat issued from your enterprise or one of its organizations. A developer signed in with a personal account, or a seat from somewhere else, doesn’t get your settings. I couldn’t find what happens then in the docs, so test that case before you trust the numbers.&lt;/li&gt;
  &lt;li&gt;&lt;strong&gt;You need somewhere to host the file.&lt;/strong&gt; If you only use GitHub to hand out &lt;em&gt;Copilot Business&lt;/em&gt; seats and have no organizations, you need one user with a GitHub Enterprise license to create an organization and the &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;.github-private&lt;/code&gt; repo. The MDM route needs none of that.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;One more thing if you set telemetry locally in VS Code instead: the agent harness reads &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;chat.agentHost.otel.*&lt;/code&gt;, and &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;enabled&lt;/code&gt; defaults to &lt;strong&gt;false&lt;/strong&gt;. Reload the window after changing it.&lt;/p&gt;

&lt;hr /&gt;

&lt;h3 id=&quot;where-this-leaves-things&quot;&gt;Where this leaves things&lt;/h3&gt;

&lt;p&gt;A small allowlist, content deleted and identity hashed in a collector you control, managed settings to ship the config, one minimal dashboard, and a re-check after every harness update.&lt;/p&gt;

&lt;p&gt;Don’t start with every signal. Write down five questions you’d actually act on, and collect only what answers them. The &lt;strong&gt;&lt;a href=&quot;https://github.com/tinestaric/harness-telemetry&quot;&gt;sample repo&lt;/a&gt;&lt;/strong&gt; has the collector, settings, queries and dashboards to start from.&lt;/p&gt;

</description>
        <pubDate>Sun, 18 Oct 2026 10:00:00 +0200</pubDate>
        <link>https://tine.staric.net/blog/2026/harness-telemetry/</link>
        <guid isPermaLink="true">https://tine.staric.net/blog/2026/harness-telemetry/</guid>
        
        
        <category>ai</category>
        
      </item>
    
      <item>
        <title>The Copilot harness is now VS Code&apos;s default</title>
        <description>&lt;p&gt;One thing that annoyed me for a while: &lt;strong&gt;&lt;em&gt;GitHub Copilot CLI&lt;/em&gt;&lt;/strong&gt;, the &lt;em&gt;VS Code Agents window&lt;/em&gt; and the &lt;em&gt;GitHub Copilot app&lt;/em&gt; all shared the same Copilot runtime, but the one most of us use every day, &lt;em&gt;Copilot Chat in VS Code&lt;/em&gt;, ran a VS Code-specific flavor of it. In my experience that made it harder to switch between tools, and some features landed elsewhere first.&lt;/p&gt;

&lt;p&gt;With &lt;strong&gt;VS Code 1.142&lt;/strong&gt;, that’s over. The &lt;em&gt;Copilot harness&lt;/em&gt; is becoming the default for new agent sessions.&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/copilotharness/harness-picker.png&quot; alt=&quot;The harness picker in chat: Copilot, Local, Cloud, Claude, Codex&quot; /&gt;&lt;/p&gt;

&lt;p&gt;The tooltip on &lt;em&gt;Copilot&lt;/em&gt; says it well: the agent experience shared consistently across GitHub Copilot products.&lt;/p&gt;

&lt;p&gt;So what does that actually get you, and what do you have to do about it?&lt;/p&gt;

&lt;hr /&gt;

&lt;h3 id=&quot;what-changes&quot;&gt;What changes&lt;/h3&gt;

&lt;p&gt;I won’t claim a huge day-one benefit. I haven’t found any benchmarks, and Microsoft doesn’t claim any either. What you get is structural:&lt;/p&gt;

&lt;ul&gt;
  &lt;li&gt;&lt;strong&gt;One behavior&lt;/strong&gt; across VS Code, the CLI, the Agents window and the Copilot app, so features stop landing in one place first.&lt;/li&gt;
  &lt;li&gt;&lt;strong&gt;Sessions run in the Agent Host&lt;/strong&gt;, a separate process from the editor window. They can keep going in the background and be shared between the editor and the Agents window, so you can switch surfaces without copying a chat over.&lt;/li&gt;
  &lt;li&gt;&lt;strong&gt;VS Code stops falling behind&lt;/strong&gt; the other Copilot surfaces.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;em&gt;Local&lt;/em&gt; is still in the picker if you want the old behavior, and it still reads your old locations.&lt;/p&gt;

&lt;hr /&gt;

&lt;h3 id=&quot;the-homework&quot;&gt;The homework&lt;/h3&gt;

&lt;p&gt;Some of your personal setup lives in places the new harness doesn’t read. The first time you open chat, you get this:&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/copilotharness/migration-notification.png&quot; alt=&quot;Notification in chat: 9 user customizations need an update to keep working, with a Review Migrations link&quot; /&gt;&lt;/p&gt;

&lt;p&gt;It doesn’t migrate anything silently, it takes you to a review list. You can also get there yourself through &lt;strong&gt;Chat: Open Customizations&lt;/strong&gt;, then &lt;strong&gt;Migrations&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/copilotharness/migrations-page.png&quot; alt=&quot;The Migrations page, listing prompt files to convert to skills and MCP servers to move&quot; /&gt;&lt;/p&gt;

&lt;p&gt;Each group has its own &lt;em&gt;Migrate&lt;/em&gt; and &lt;em&gt;Ignore&lt;/em&gt;. &lt;em&gt;Migrate&lt;/em&gt; lets VS Code move the files itself. &lt;em&gt;Migrate with Agent&lt;/em&gt; does it as a guided session in chat, which uses credits.&lt;/p&gt;

&lt;p&gt;Nothing in the repo moves. Instructions, &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;AGENTS.md&lt;/code&gt; and agents under &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;.github&lt;/code&gt; stay where they are. What moves is your &lt;em&gt;personal&lt;/em&gt; stuff, into &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;~/.copilot&lt;/code&gt;:&lt;/p&gt;

&lt;table&gt;
  &lt;thead&gt;
    &lt;tr&gt;
      &lt;th&gt;What&lt;/th&gt;
      &lt;th&gt;Where it goes&lt;/th&gt;
    &lt;/tr&gt;
  &lt;/thead&gt;
  &lt;tbody&gt;
    &lt;tr&gt;
      &lt;td&gt;Personal instructions&lt;/td&gt;
      &lt;td&gt;&lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;~/.copilot/copilot-instructions.md&lt;/code&gt; and &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;~/.copilot/instructions/&lt;/code&gt;&lt;/td&gt;
    &lt;/tr&gt;
    &lt;tr&gt;
      &lt;td&gt;Personal agents&lt;/td&gt;
      &lt;td&gt;&lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;~/.copilot/agents/&lt;/code&gt;&lt;/td&gt;
    &lt;/tr&gt;
    &lt;tr&gt;
      &lt;td&gt;Prompt files (&lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;*.prompt.md&lt;/code&gt;)&lt;/td&gt;
      &lt;td&gt;converted to skills, e.g. in &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;~/.copilot/skills&lt;/code&gt;&lt;/td&gt;
    &lt;/tr&gt;
    &lt;tr&gt;
      &lt;td&gt;User MCP servers&lt;/td&gt;
      &lt;td&gt;&lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;~/.copilot/mcp-config.json&lt;/code&gt;&lt;/td&gt;
    &lt;/tr&gt;
    &lt;tr&gt;
      &lt;td&gt;Workspace MCP servers&lt;/td&gt;
      &lt;td&gt;&lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;.mcp.json&lt;/code&gt; in the repo&lt;/td&gt;
    &lt;/tr&gt;
  &lt;/tbody&gt;
&lt;/table&gt;

&lt;p&gt;Two things in my own list are worth knowing before you click anything.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Prompt files become skills, and lose some of their powers.&lt;/strong&gt; The &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;agent&lt;/code&gt;, &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;model&lt;/code&gt; and &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;tools&lt;/code&gt; headers get dropped, and the converted skills are set to &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;disable-model-invocation: true&lt;/code&gt;. The agent won’t pick them up by itself anymore, you invoke them manually with &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;/name&lt;/code&gt;. If you had a prompt file that was supposed to be triggered automatically, that part is gone.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Migrating MCP servers installed from the registry stops their auto-updates.&lt;/strong&gt; The screen warns you about it in orange, right under the server. I migrated mine anyway, but now I’m the one who has to remember to update them.&lt;/p&gt;

&lt;hr /&gt;

&lt;h3 id=&quot;how-id-do-it&quot;&gt;How I’d do it&lt;/h3&gt;

&lt;p&gt;Commit first, migrate, then test in a fresh Copilot session before touching anything else. The migrated files are copies, not moves, so the originals stay where they were. Delete those last, once you know the new setup works, and don’t leave both active in the meantime.&lt;/p&gt;

&lt;p&gt;Also keep in mind that this is a one-time move per machine. I haven’t checked whether &lt;em&gt;Settings Sync&lt;/em&gt; carries &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;~/.copilot&lt;/code&gt; over, so don’t assume your personal setup follows you to the next laptop.&lt;/p&gt;

&lt;hr /&gt;

&lt;p&gt;If you’ve already gone through the migration, I’m curious what broke for you. Mine was mostly prompt files and MCP servers, so I got off easy.&lt;/p&gt;
</description>
        <pubDate>Fri, 16 Oct 2026 10:00:00 +0200</pubDate>
        <link>https://tine.staric.net/blog/2026/copilot-harness-default/</link>
        <guid isPermaLink="true">https://tine.staric.net/blog/2026/copilot-harness-default/</guid>
        
        
        <category>al</category>
        
      </item>
    
      <item>
        <title>Claude Code: Tiny Tips</title>
        <description>&lt;p&gt;When we move devs to &lt;em&gt;Claude Code&lt;/em&gt;, I usually give them a quick “here’s what’s good to know about this tool” rundown. After a few rounds of that, a handful of commands keep coming up as &lt;em&gt;“huh, didn’t know that.”&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;Here’s my go-to list. Note: I use the terminal, not the VS Code extension, so not everything below is accessible from there as far as I know.&lt;/p&gt;

&lt;h3 id=&quot;map-a-shortcut-for-new-terminals&quot;&gt;Map a shortcut for new terminals&lt;/h3&gt;

&lt;p&gt;In VS Code, map the &lt;strong&gt;“Terminal: Create Terminal in Editor Area”&lt;/strong&gt; action to a keyboard shortcut so you can spin up a new &lt;em&gt;Claude Code&lt;/em&gt; session without touching the mouse. I have mine on &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;Ctrl + Alt + T&lt;/code&gt;.&lt;/p&gt;

&lt;h3 id=&quot;stop-the-terminal-from-swallowing-your-shortcuts&quot;&gt;Stop the terminal from swallowing your shortcuts&lt;/h3&gt;

&lt;p&gt;With a terminal focused, &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;Ctrl + B&lt;/code&gt; should toggle VS Code’s sidebar, but the shell running inside the terminal grabs the keypress first and VS Code never sees it. Add &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;workbench.action.toggleSidebarVisibility&lt;/code&gt; to &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;terminal.integrated.commandsToSkipShell&lt;/code&gt; in your &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;settings.json&lt;/code&gt; and VS Code handles the shortcut itself instead of forwarding it to the shell:&lt;/p&gt;

&lt;div class=&quot;language-json highlighter-rouge&quot;&gt;&lt;div class=&quot;highlight&quot;&gt;&lt;pre class=&quot;highlight&quot;&gt;&lt;code&gt;&lt;span class=&quot;nl&quot;&gt;&quot;terminal.integrated.commandsToSkipShell&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;p&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
    &lt;/span&gt;&lt;span class=&quot;s2&quot;&gt;&quot;workbench.action.zoomIn&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
    &lt;/span&gt;&lt;span class=&quot;s2&quot;&gt;&quot;workbench.action.zoomOut&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
    &lt;/span&gt;&lt;span class=&quot;s2&quot;&gt;&quot;workbench.action.toggleSidebarVisibility&quot;&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/div&gt;&lt;/div&gt;

&lt;h3 id=&quot;tui-fullscreen&quot;&gt;&lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;/tui fullscreen&lt;/code&gt;&lt;/h3&gt;

&lt;p&gt;Cleans up the terminal GUI and removes the flicker you get on every redraw.&lt;/p&gt;

&lt;h3 id=&quot;porting-agents-isnt-a-copy-paste&quot;&gt;Porting agents isn’t a copy-paste&lt;/h3&gt;

&lt;p&gt;You can’t port agents directly to &lt;em&gt;Claude Code&lt;/em&gt; because the tool names are different from &lt;em&gt;Copilot&lt;/em&gt;’s. Ask the AI for help translating them, I’ve got a whole separate post on this one alone: &lt;strong&gt;&lt;em&gt;&lt;a href=&quot;/blog/2026/copilot-to-claude-code/&quot;&gt;“Copilot is agent-first, Claude Code is skill-first”&lt;/a&gt;&lt;/em&gt;&lt;/strong&gt;.&lt;/p&gt;

&lt;h3 id=&quot;doctor&quot;&gt;&lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;/doctor&lt;/code&gt;&lt;/h3&gt;

&lt;p&gt;Lets the AI analyze your agents, skills, and instructions. Quite useful right after you’ve ported something over and want to know what can be improved.&lt;/p&gt;

&lt;h3 id=&quot;model-opusplan&quot;&gt;&lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;/model opusplan&lt;/code&gt;&lt;/h3&gt;

&lt;p&gt;&lt;em&gt;Opus&lt;/em&gt; for planning, &lt;em&gt;Sonnet&lt;/em&gt; for execution, switched automatically as you move between the two. This option isn’t selectable through the UI, you have to set it with the command.&lt;/p&gt;

&lt;h3 id=&quot;advisor&quot;&gt;&lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;/advisor&lt;/code&gt;&lt;/h3&gt;

&lt;p&gt;When &lt;em&gt;Sonnet&lt;/em&gt; gets stuck, it can reach out to &lt;em&gt;Opus&lt;/em&gt; or &lt;em&gt;Fable&lt;/em&gt; on its own for a second opinion.&lt;/p&gt;

&lt;h3 id=&quot;rewind&quot;&gt;&lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;/rewind&lt;/code&gt;&lt;/h3&gt;

&lt;p&gt;Undoes changes and resets the state and conversation back to a previous prompt. Cheaper than starting a whole new session when a change went sideways.&lt;/p&gt;

&lt;h3 id=&quot;left-arrow-key&quot;&gt;Left arrow key&lt;/h3&gt;

&lt;p&gt;Gets you to the overview of all sessions. I prefer separate terminal windows per session myself, but some people like having this single view instead.&lt;/p&gt;

&lt;hr /&gt;

&lt;p&gt;Do you have any tiny tips of your own that help you work with &lt;em&gt;Claude Code&lt;/em&gt;? I’d genuinely like to hear them.&lt;/p&gt;

</description>
        <pubDate>Mon, 07 Sep 2026 10:00:00 +0200</pubDate>
        <link>https://tine.staric.net/blog/2026/claude-code-tiny-tips/</link>
        <guid isPermaLink="true">https://tine.staric.net/blog/2026/claude-code-tiny-tips/</guid>
        
        
        <category>al</category>
        
      </item>
    
      <item>
        <title>Context is a voting chamber</title>
        <description>&lt;p&gt;Should you have a long agent session that has all the context you need, or should you aim for short and focused sessions? I think we all already have the gut feeling that short is better and yet we never want to abandon a long session. We usually like the first response from an agent, but the 10th iteration, where it’s running in loops and still making the same mistakes is where it starts to get annoying.&lt;/p&gt;

&lt;p&gt;The reason people usually quote against long sessions is “it gets expensive”. Partially true. The models do bill for input and output tokens, and the more context you input, the more expensive the call is. However the main argument against long sessions isn’t economics, but something else. To understand why long sessions are both okay and not okay, we need to understand two mechanisms that are at play in the background: caching and attention.&lt;/p&gt;

&lt;hr /&gt;

&lt;h3 id=&quot;caching-the-reason-long-sessions-arent-as-expensive-as-you-think&quot;&gt;Caching: the reason long sessions aren’t as expensive as you think&lt;/h3&gt;

&lt;p&gt;When you look at any diagram of “how LLMs work”, you’ll see these scary looking images: a token going in one side, climbing through eighty-some stacked layers, something coming out the other end.&lt;/p&gt;

&lt;p style=&quot;text-align: center;&quot;&gt;
  &lt;img src=&quot;/images/outvoted/token-journey.svg&quot; alt=&quot;One token climbing through the layers of a model, meeting attention heads and neurons at every layer, emerging as an output vector&quot; style=&quot;max-width: 100%;&quot; /&gt;
&lt;/p&gt;

&lt;p&gt;For today, forget all of that. The only thing you need to know is that every token you send gets turned into a set of numbers, and those numbers are what the model actually works with to figure out what comes next.&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/outvoted/tokens-to-vectors.svg&quot; alt=&quot;A sentence split into four tokens, each turning into its own key/value vector of numbers, which get stored for reuse&quot; /&gt;&lt;/p&gt;

&lt;p&gt;The model bills for every one of those tokens, in and out. But underneath, there are two more meters running that you usually never look at, cache write and cache read. Here’s what all of that costs on Opus 5, per million tokens:&lt;/p&gt;

&lt;ul&gt;
  &lt;li&gt;&lt;strong&gt;Input:&lt;/strong&gt; $5.00, for every token the model has to actually work through for the first time&lt;/li&gt;
  &lt;li&gt;&lt;strong&gt;Output:&lt;/strong&gt; $25.00, for every token it generates back (this includes thinking tokens)&lt;/li&gt;
  &lt;li&gt;&lt;strong&gt;Cache write (5-minute):&lt;/strong&gt; $6.25, for storing the current prefix so it can be reused later&lt;/li&gt;
  &lt;li&gt;&lt;strong&gt;Cache write (1-hour):&lt;/strong&gt; $10.00, for storing it on the longer-lived tier instead&lt;/li&gt;
  &lt;li&gt;&lt;strong&gt;Cache read:&lt;/strong&gt; $0.50, for reusing something that’s already sitting in the cache&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;That last number is the one that matters. If the start of your conversation hasn’t changed since your last message, the vectors behind it haven’t changed either, so nothing forces the model to recompute them. It just reads what’s already stored, at a tenth of the input price. That’s prompt caching. Most of a long conversation genuinely does bill at that cache-read rate, because most of it is being read, not reprocessed.&lt;/p&gt;

&lt;p&gt;Except it isn’t only your prompts that get billed this way. An agent doesn’t wait for you to type before it acts. Every tool call it makes, every time it decides what to do next, is its own request to the model, and each one rebills the entire accumulated prefix at that same read rate, not just whatever’s new since the last call.&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/outvoted/reread-cards.svg&quot; alt=&quot;Calls 1, 2, 5, and 10 of an agent fixing a rounding bug in Sales-Post.Codeunit.al, each one rereading everything before it, ending at a 23,000-token conversation&quot; /&gt;&lt;/p&gt;

&lt;p&gt;Ten steps into a session whose final context sits at 23k tokens, you’ve billed something like 119k cache-read tokens across those ten calls. The context grew linearly. The bill grew quadratically.&lt;/p&gt;

&lt;p&gt;But let’s look at the number again. 119k tokens at a tenth of input price is roughly the same bill as 12k tokens at full rate. For ten steps of agent work, that’s nothing. Quadratic growth is real but also boringly affordable.&lt;/p&gt;

&lt;p&gt;Do remember though that cache isn’t forever. Anthropic offers two lifetimes, five minutes or an hour, and which one you’re on depends on how you’re working. &lt;strong&gt;&lt;em&gt;Claude Code&lt;/em&gt;&lt;/strong&gt; defaults its main conversation to the one-hour cache, though subagents still run on the five-minute one. Go through the API directly on usage-based billing, and you default to five minutes too. Either way it refreshes on every hit, so a session you keep working stays warm indefinitely for free. Walk away longer than that, for lunch or a meeting, and you’re paying for the whole prefix again.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Side note for the &lt;strong&gt;GitHub Copilot&lt;/strong&gt; crowd: if you’re on an OpenAI model, the default cache is on a similar clock, five to ten minutes of inactivity, up to an hour in some cases. For supported OpenAI models, VS Code now sets &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;prompt_cache_retention: &quot;24h&quot;&lt;/code&gt; on your behalf, which trades a slower cache tier for surviving a lunch break intact.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;So: while the cache is warm, long sessions are genuinely manageable. Carry on.&lt;/p&gt;

&lt;hr /&gt;

&lt;h3 id=&quot;attention-the-reason-long-sessions-are-still-a-bad-idea&quot;&gt;Attention: the reason long sessions are still a bad idea&lt;/h3&gt;

&lt;p&gt;&lt;strong&gt;Cost, then, isn’t the problem. Attention is&lt;/strong&gt;, but there’s no meter for this one.&lt;/p&gt;

&lt;p&gt;Remember that diagram from earlier, the layers the token climbed through? The brown boxes in there are attention heads, dozens of them per layer, each one deciding for itself how much of the surrounding context matters right now. That’s the model’s actual mechanism for attention. Every one of those heads divides one unit of attention across everything in its context, every single time it produces the next token, and none of them can opt out of spending it. That’s what softmax does, for anyone who wants to look it up later.&lt;/p&gt;

&lt;p&gt;What that means in practice: every voter’s weight depends on one thing: how much it looks like the answer to the question being asked right now. A token that clearly isn’t the answer gets almost nothing, and two hundred of them together still get almost nothing. What splits the vote is a handful of tokens that each look like they might be it. And a long session is made of exactly those: the same file read three times, the same plan revised twice, the same error pasted back from four tool calls. Nothing about the relevant token changed. It’s sharing the room with its own near-copies.&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/outvoted/dilution-bars.svg&quot; alt=&quot;The same relevant token holding steady against 200 off-topic tokens, then splitting its share three ways against near-duplicates of itself&quot; /&gt;&lt;/p&gt;

&lt;p&gt;Position makes this worse before it makes it better. Models attend well to the end of the context, because recent tokens are usually the most predictive of what comes next, and reasonably well to the very start. The middle just sits there. That’s measured on how well models actually retrieve things, not on the attention weights directly, but it shows up consistently. Whatever you told the agent at the top of a long session, your actual objective, is now buried in exactly the part of context that gets the least attention of anyone in the room.&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/outvoted/position-curve.svg&quot; alt=&quot;Attention drawn as a curve over the context window, high at both ends, dipping in the middle where your original objective sits&quot; /&gt;&lt;/p&gt;

&lt;hr /&gt;

&lt;h3 id=&quot;the-voting-chamber&quot;&gt;The voting chamber&lt;/h3&gt;

&lt;p&gt;Here’s the way I’ve started thinking about it. Context isn’t a desk where old information sits quietly until you need it. &lt;strong&gt;It’s a voting chamber.&lt;/strong&gt; Everything inside votes on the next token, every time, weighted by relevance, always summing to one. Nobody sits out completely, most just end up with almost nothing to say.&lt;/p&gt;

&lt;p&gt;Three situations this explains better than “the agent forgot”:&lt;/p&gt;

&lt;p&gt;You read a file, edit it, read it again, edit it again, read it a third time. Context now holds three versions of that file. Three voters, all confidently answering “what’s actually in this file”, all convinced they’re right, because at the time each was written, each was. The current version has no special standing. It’s one of three votes, slightly favoured by recency.&lt;/p&gt;

&lt;p&gt;You ask the agent for an email, and it writes one. You don’t like the subject line, so you fix that. The close feels off, so you fix that too. A few rounds in, everything still leans back toward that first draft’s structure and tone no matter what you correct, because the first email is still sitting in context, complete and confident, the only voter with a full answer to “what does an email look like around here.” Every correction you send is one voice arguing against a voter that showed up with an entire draft behind it. It doesn’t stand a chance. Scrap the thread instead, and rewrite the original prompt with everything you’ve learned since folded in from the start.&lt;/p&gt;

&lt;p&gt;An approach you told the agent to abandon is still in the room. “Drop it” wins for a few turns, while it’s freshest. Then both get buried and the approach creeps back, because every vote is a vote on what to write next, and “drop it” proposes nothing. The only voter holding actual lines is the approach you abandoned, so those get written again, slightly disguised.&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/outvoted/chamber-3-abandoned-bars.svg&quot; alt=&quot;Right after &amp;quot;drop it&amp;quot;, the retraction is loud and the approach fades; a few turns later the retraction has faded out while the approach still holds actual lines to write&quot; /&gt;&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/outvoted/chamber-2-vignettes.svg&quot; alt=&quot;Three chamber panels, one per scenario: three file versions, one email draft against its two corrections, one approach description against its retraction, all voting for the next token&quot; /&gt;&lt;/p&gt;

&lt;p&gt;The obvious instinct is that the only way to clear the room is to end the session and start over. That’s not quite right. Your harness already retires some voters on its own, &lt;em&gt;Claude Code&lt;/em&gt; keeps one running todo list instead of ten stale copies, plan mode edits the plan itself instead of appending revision after revision, and old tool results age out or get replaced with placeholders as the session goes on, exactly the retirement mechanism the read-edit-read example above needs. You do have the tool, but it’s the harness that decides which voter goes and when, not you.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;You cannot mark something as irrelevant from inside the conversation.&lt;/strong&gt; Telling the agent “ignore that email format” doesn’t remove a voter. It adds one. Now there’s a fourth voice in the room, arguing against the third, and the third hasn’t gone anywhere. Whatever gets retired automatically retires on the harness’s schedule, not yours. Anything else still means summarizing what matters yourself, and starting over.&lt;/p&gt;

&lt;hr /&gt;

&lt;h3 id=&quot;what-to-do-instead&quot;&gt;What to do instead&lt;/h3&gt;

&lt;p&gt;What the harness can’t automate is judgment: which of your abandoned approaches are worth a full paragraph and which aren’t, when a conversation has drifted so far from its actual goal that no more correcting will fix it. That’s still yours to do. Here’s what to do about it.&lt;/p&gt;

&lt;p&gt;Write plans, and hand off between them, more often than feels necessary. Break the work into sections, give each section to a fresh session, have it write a handoff for the next one, and start that one clean. A handoff should preserve the eliminated options, not the narrative. “We tried X, it broke Y, so we went with Z instead” is the expensive, hard-won part. Drop it, and the next session just tries X again.&lt;/p&gt;

&lt;p&gt;Use subagents for anything that involves reading a lot. &lt;strong&gt;&lt;em&gt;&lt;a href=&quot;/blog/2026/copilot-fancy-autocomplete-vol3/&quot;&gt;Vol.3&lt;/a&gt;&lt;/em&gt;&lt;/strong&gt; has the example: three subagents digging through the base app, the docs, and a specialist in parallel, each one burning its own context on the reading, and handing the main session back a condensed report instead of the transcript. Same benefit as a handoff, just paid before the pollution reaches the main conversation instead of cleaned up afterwards.&lt;/p&gt;

&lt;p&gt;Resuming an old session runs into the same problem, and &lt;em&gt;Claude Code&lt;/em&gt; will actually warn you about it now:&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/outvoted/resume-from-summary.png&quot; alt=&quot;Claude Code prompting to resume a 6-day-old, 644.6k-token session from a summary instead of in full&quot; /&gt;&lt;/p&gt;

&lt;p&gt;Even at the cache-read rate, 644.6k tokens is the equivalent of about 64k tokens at full input price, and that’s not a one-time cost, it’s what every single agent decision from that point on drags along with it. Don’t resume old sessions in full if you can help it. If you do resume one, resume from a summary.&lt;/p&gt;

&lt;hr /&gt;

&lt;h3 id=&quot;the-trade-youre-making&quot;&gt;The trade you’re making&lt;/h3&gt;

&lt;p&gt;The two mechanisms in this post want opposite things. Caching rewards never touching what came before, append only, everything stays warm. Attention rewards removing what’s gone stale, which means rewriting the prefix, and that burns the cache.&lt;/p&gt;

&lt;p&gt;So every handoff is a deliberate cache burn, and it’s worth it: you’re spending real money to buy back the one resource in this whole system that has no meter attached to it.&lt;/p&gt;

&lt;p&gt;And it’s cheaper than “deliberate cache burn” makes it sound. A handoff costs one fresh cache write at 1.25 times the base rate, but on a small prefix, because the new session starts lean. That’s a rounding error next to what staying unfocused would have cost you instead.&lt;/p&gt;

</description>
        <pubDate>Mon, 31 Aug 2026 10:00:00 +0200</pubDate>
        <link>https://tine.staric.net/blog/2026/outvoted/</link>
        <guid isPermaLink="true">https://tine.staric.net/blog/2026/outvoted/</guid>
        
        
        <category>al</category>
        
      </item>
    
      <item>
        <title>Copilot is agent-first, Claude Code is skill-first</title>
        <description>&lt;p&gt;I love &lt;strong&gt;&lt;em&gt;GitHub Copilot&lt;/em&gt;&lt;/strong&gt;. I always have. But I also started using &lt;strong&gt;&lt;em&gt;Claude Code&lt;/em&gt;&lt;/strong&gt; as my main driver a few months ago, and I figured the transition would be smooth. It mostly was, for me. Then we started trying &lt;em&gt;Claude Code&lt;/em&gt; out for part of our organization, and I found out fast that it’s not really a one-to-one switch. Not when you’ve already built agents and skills specifically for Copilot.&lt;/p&gt;

&lt;p&gt;The most common question I get from people making that switch is &lt;em&gt;“how do I switch to my custom agents now?”&lt;/em&gt; And the answer is: &lt;strong&gt;you don’t&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/copilottoclaudecode/copilot-agent-dropdown.png&quot; alt=&quot;Copilot&apos;s agent dropdown, a list of custom agents to step into&quot; /&gt;&lt;/p&gt;

&lt;p&gt;That answer isn’t useful on its own though. So here’s what I’ve learned about carrying an agentic workflow from &lt;em&gt;Copilot&lt;/em&gt; over to &lt;em&gt;Claude Code&lt;/em&gt;, and why it made me realize that Copilot is “agent-first” and Claude Code is “skill-first”.&lt;/p&gt;

&lt;p&gt;The first instinct, mine included, is that &lt;em&gt;agent&lt;/em&gt; becomes &lt;em&gt;agent&lt;/em&gt;. Same word, same idea of a markdown file with frontmatter describing a persona and a set of tools, surely it just carries over. It doesn’t, not fully.&lt;/p&gt;

&lt;p&gt;Every custom agent in &lt;em&gt;Claude Code&lt;/em&gt; is &lt;strong&gt;a subagent&lt;/strong&gt;. There’s no other kind. Whatever you define, it forks off a clean, empty context window, does its work in there, and hands your main session back a condensed report. Nothing it read or tried comes back with it, only the summary.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Copilot&lt;/em&gt; has two different kinds of custom agents. The ones you spin up through &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;runSubagent&lt;/code&gt;, the ones behind &lt;strong&gt;&lt;em&gt;&lt;a href=&quot;/blog/2026/copilot-fancy-autocomplete-vol3/&quot;&gt;vol.3’s&lt;/a&gt;&lt;/em&gt;&lt;/strong&gt; parallel research example, already fork and report back exactly like a &lt;em&gt;Claude Code&lt;/em&gt; subagent does. Those transfer directly.&lt;/p&gt;

&lt;p&gt;The ones you step into from the dropdown are the other kind, and they don’t behave the same way at all. Stepping into one prepends its whole body, persona and instructions and all, to the conversation you’re already having. It’s still your conversation, just handed a new voice, nothing forks, nothing comes back as a summary because nothing ever left. That’s the half of your scaffold that doesn’t have an agent waiting for it on the Claude Code side.&lt;/p&gt;

&lt;p&gt;This one actually maps closer to a &lt;strong&gt;skill&lt;/strong&gt; than a subagent.&lt;/p&gt;

&lt;hr /&gt;

&lt;h3 id=&quot;why-claude-code-feels-skill-shaped&quot;&gt;Why Claude Code feels skill-shaped&lt;/h3&gt;

&lt;p&gt;&lt;em&gt;Copilot&lt;/em&gt; is agent-first: you pick who you’re talking to before you say anything. &lt;em&gt;Claude Code&lt;/em&gt; is skill-first, because there’s only ever one &lt;em&gt;you’re talking to&lt;/em&gt;, the same agent from the moment you open the terminal to the moment you close it. You don’t instantiate a persona. You tell the one you already have what it needs to know for this particular task.&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/copilottoclaudecode/claude-code-nothing-to-select.png&quot; alt=&quot;Claude Code&apos;s startup screen, nothing to select&quot; /&gt;&lt;/p&gt;

&lt;p&gt;In &lt;em&gt;Copilot&lt;/em&gt;, selecting an agent from the dropdown was the act of giving your conversation context: this persona, these instructions, this scoped set of tools. In &lt;em&gt;Claude Code&lt;/em&gt;, that same act is invoking a skill. What used to be &lt;em&gt;“step into Backlog Reviewer, type start with task 1123”&lt;/em&gt; becomes &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;/start-feature 1123&lt;/code&gt;.&lt;/p&gt;

&lt;hr /&gt;

&lt;h3 id=&quot;dont-over-rotate&quot;&gt;Don’t over-rotate&lt;/h3&gt;

&lt;p&gt;If you stop there and turn every single Copilot agent into a skill, you’ll have thrown away context isolation on the way.&lt;/p&gt;

&lt;p&gt;Here’s a test you can run against your own memory instead of a rule to memorize. Think about the last few times you stepped into a Copilot agent. When it finished, did you keep working in that same chat? Or did you take what it gave you, a plan, a summary, a list of files, and go start something else with it?&lt;/p&gt;

&lt;p&gt;If you routinely carried the output somewhere else, you were doing context isolation by hand, one copy-paste at a time. That agent wanted to be a subagent all along.&lt;/p&gt;

&lt;hr /&gt;

&lt;h3 id=&quot;so-is-it-a-skill-or-a-subagent&quot;&gt;So is it a skill or a subagent?&lt;/h3&gt;

&lt;p&gt;Here’s how I decide:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Do you need to type it?&lt;/strong&gt; If you want a &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;/something&lt;/code&gt; you can fire on demand, it’s a skill. You can ask Claude Code to use a particular subagent, but you can’t invoke one by name the way you’d pick an agent from a dropdown.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Where does the work belong?&lt;/strong&gt; In the conversation you’re already having (skill), or in a clean window that reports back when it’s done? (subagent)&lt;/p&gt;

&lt;p&gt;The two do overlap a bit. If you want to type it and you want it isolated, a skill can be told to fork (&lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;context: fork&lt;/code&gt;), but that’s a post of its own and you don’t need it to do the sort.&lt;/p&gt;

&lt;hr /&gt;

&lt;h3 id=&quot;the-tools-list-breaks&quot;&gt;The tools list breaks&lt;/h3&gt;

&lt;p&gt;When you move agents to Claude Code, two things happen.&lt;/p&gt;

&lt;p&gt;The tools list breaks for a shallow reason first: the names are just different. A &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;tools:&lt;/code&gt; array full of Copilot’s &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;edit&lt;/code&gt; and &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;search&lt;/code&gt; means nothing to &lt;em&gt;Claude Code&lt;/em&gt;, which wants &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;Edit&lt;/code&gt; and &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;Glob&lt;/code&gt;. I ran into this shipping a plugin across both ecosystems, the fix was listing both platforms’ names side by side in one array, each tool only recognizing its own, covered in &lt;strong&gt;&lt;em&gt;&lt;a href=&quot;/blog/2026/agent-plugins/&quot;&gt;the plugins post&lt;/a&gt;&lt;/em&gt;&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/copilottoclaudecode/unknown-tool-warning.png&quot; alt=&quot;VS Code flagging Read as an unknown tool, Copilot doesn&apos;t recognize Claude Code&apos;s name for it&quot; /&gt;&lt;/p&gt;

&lt;p&gt;But there’s a second, bigger break underneath that one. In &lt;em&gt;Copilot&lt;/em&gt;, restricting tools was a property of the agent you stepped into. Step into the read-only reviewer, and for as long as you’re in that chat, editing isn’t on the table. In &lt;em&gt;Claude Code&lt;/em&gt;, a skill loaded inline can’t restrict anything, because it’s just text your main agent reads and can weigh against everything else in the conversation. Your main session holds every tool and every MCP server you’ve configured, and loading a skill only adds words on top, nothing gets taken away.&lt;/p&gt;

&lt;p&gt;In &lt;em&gt;Copilot&lt;/em&gt; that guarantee lived in the agent you stepped into. In &lt;em&gt;Claude Code&lt;/em&gt; it has to live in the thing you dispatch to. So if &lt;em&gt;“this must never touch a file”&lt;/em&gt; is critical, &lt;strong&gt;don’t put it in a skill, put it in a subagent, where &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;Edit&lt;/code&gt; isn’t an option&lt;/strong&gt;.&lt;/p&gt;

&lt;hr /&gt;

&lt;h3 id=&quot;can-you-really-have-one-plugin-for-both-ecosystems&quot;&gt;Can you really have one plugin for both ecosystems?&lt;/h3&gt;

&lt;p&gt;Not as cleanly as I made it sound in &lt;strong&gt;&lt;em&gt;&lt;a href=&quot;/blog/2026/agent-plugins/&quot;&gt;the plugins post&lt;/a&gt;&lt;/em&gt;&lt;/strong&gt;, one shared &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;agents/&lt;/code&gt; folder, same file read by both ecosystems. That’s still true for where the file lives. It’s less true based on what’s inside, because a &lt;em&gt;Copilot&lt;/em&gt; dropdown agent and a &lt;em&gt;Claude Code&lt;/em&gt; subagent are playing two different roles. In practice I’ve ended up diverging most of mine per platform rather than sharing them, so it’s worth deciding upfront whether you actually want one shared set of agents, or you’re fine maintaining two. All skills I had on the Copilot side however are still the same skills on the Claude Code side.&lt;/p&gt;

&lt;hr /&gt;

&lt;p&gt;Don’t translate your Copilot agent folder file-for-file. Sort each agent by what it actually did, isolated fork or same-conversation persona, not by what kind of file it lives in, and let the skill/subagent split in &lt;em&gt;Claude Code&lt;/em&gt; fall out of that sort. But don’t worry about your skills, they carry over just fine.&lt;/p&gt;

</description>
        <pubDate>Fri, 21 Aug 2026 10:00:00 +0200</pubDate>
        <link>https://tine.staric.net/blog/2026/copilot-to-claude-code/</link>
        <guid isPermaLink="true">https://tine.staric.net/blog/2026/copilot-to-claude-code/</guid>
        
        
        <category>al</category>
        
      </item>
    
      <item>
        <title>Distributing Agents with Plugins</title>
        <description>&lt;p&gt;The post &lt;strong&gt;&lt;em&gt;&lt;a href=&quot;/blog/2026/copilot-fancy-autocomplete-vol3/&quot;&gt;Copilot is just a fancy autocomplete vol.3&lt;/a&gt;&lt;/em&gt;&lt;/strong&gt; ended with “go build yourself one agent, just one, it doesn’t need to be the whole suite.” Turns out that’s the easy part. The question I actually get, almost every time I run a session on building custom agents, is the one that post doesn’t answer at all: &lt;strong&gt;okay, I built one, now how does the rest of my team get it?&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;I’ve been sitting on this question since &lt;strong&gt;&lt;em&gt;EMEA 2025&lt;/em&gt;&lt;/strong&gt;, where &lt;strong&gt;&lt;em&gt;&lt;a href=&quot;https://www.linkedin.com/in/ericwauters&quot;&gt;waldo&lt;/a&gt;&lt;/em&gt;&lt;/strong&gt; told me, mid-conversation, that he’d just vibe-coded a &lt;em&gt;VS Code extension&lt;/em&gt; that drops his team’s agents straight into everyone’s editor. I thought it was a great idea. I also didn’t move an inch on it, because I was still convinced everyone should be building their own agents first. Handing someone a finished suite felt like the wrong lesson to teach this early.&lt;/p&gt;

&lt;p&gt;The push to actually revisit it came a few days ago, when waldo published &lt;strong&gt;&lt;em&gt;&lt;a href=&quot;https://waldo.be/2026/06/29/the-waldo-way/&quot;&gt;The Waldo Way&lt;/a&gt;&lt;/em&gt;&lt;/strong&gt;, the writeup of that same extension, now called the &lt;strong&gt;&lt;em&gt;iFacto Playbook&lt;/em&gt;&lt;/strong&gt;, and the doubt that came with shipping it: did removing all the setup friction also remove the learning that used to happen inside that friction? I don’t have that answer either. But reading it put the original question back in front of me, and this time I actually sat with it instead of shelving it again.&lt;/p&gt;

&lt;p&gt;A &lt;strong&gt;default set of agents&lt;/strong&gt; doesn’t have to cost you the other option. It can spark exploration for the people who’ll eventually go build their own. And it stops blocking the people who don’t have the time this sprint, whose deadline doesn’t care whether they’ve found the time to play around with agents yet.&lt;/p&gt;

&lt;p&gt;So: we need a default set. The only real question left is how you get it in front of people without it turning into “well, it works, but is this actually what I’d recommend to every partner?”&lt;/p&gt;

&lt;hr /&gt;

&lt;h3 id=&quot;nearly-missing-the-point-of-packaging&quot;&gt;Nearly missing the point of packaging&lt;/h3&gt;

&lt;p&gt;My first instinct was to do what waldo did: build the extension, ship it to the team. But “is this really the recommendation I’m going to give to every single partner?” kept nagging at me. Somebody always has to maintain it, and it’s not a lot of maintenance, but it also becomes a blocker the moment there isn’t a plug-and-play way for the next partner to get the same thing.&lt;/p&gt;

&lt;p&gt;So my mind went to a community extension instead: something generic you point at a repository, and it pulls everything down for everyone in your org. Which, funnily enough, is almost exactly what already exists. I just didn’t recognize it, because I’d already dismissed the concept once this year.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Plugins.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;They were the one piece I never fully got when I was relearning this whole space at the start of the year. Agents, skills, MCPs, hooks, those all clicked fast, they’re just files you can already share, point someone at a repo and tell them to copy them in. So I couldn’t see what a “plugin” added on top. Packaging felt like a solution to a problem I didn’t have.&lt;/p&gt;

&lt;p&gt;Turns out I just hadn’t asked the right question yet. The moment I framed it as &lt;em&gt;distribution&lt;/em&gt;, how do these files reach everyone’s editor without every developer manually copying files into the right folder, it was an instant, face-palm-obvious “oh, that’s what packaging is for.” A plugin isn’t a new way to write an agent or a skill. It’s a folder that bundles the ones you already have, hosted in a repo, that anyone can install with one click and update the same way.&lt;/p&gt;

&lt;p&gt;And the part that actually got me to build one: the exact same plugin repo works for both &lt;strong&gt;GitHub Copilot&lt;/strong&gt; and &lt;strong&gt;Claude Code&lt;/strong&gt;. One repo, two ecosystems.&lt;/p&gt;

&lt;p&gt;Here’s the install experience, from the other side, the side your teammates actually see.&lt;/p&gt;

&lt;p&gt;First, someone adds your repo as a marketplace, once, in their &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;settings.json&lt;/code&gt;:&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/agentplugins/settings-marketplace.png&quot; alt=&quot;Registering a plugin marketplace in settings.json&quot; /&gt;&lt;/p&gt;

&lt;p&gt;Then they open the &lt;em&gt;Agent Customizations&lt;/em&gt; panel from the command palette:&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/agentplugins/open-customizations-command.png&quot; alt=&quot;Opening Chat: Open Customizations from the command palette&quot; /&gt;&lt;/p&gt;

&lt;p&gt;Go to &lt;em&gt;Plugins&lt;/em&gt;, hit &lt;em&gt;Browse Marketplace&lt;/em&gt;, and there’s the plugin, ready to install:&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/agentplugins/plugins-browse-marketplace.png&quot; alt=&quot;Browsing the marketplace and finding the plugin&quot; /&gt;&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/agentplugins/plugin-install.png&quot; alt=&quot;Installing a plugin from the marketplace&quot; /&gt;&lt;/p&gt;

&lt;p&gt;One click, and every agent in that plugin shows up in the agent picker, no config, no copy-pasting &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;.md&lt;/code&gt; files into the right folder by hand:&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/agentplugins/agent-list-dropdown.png&quot; alt=&quot;All the agents from the plugin now available in the agent picker&quot; /&gt;&lt;/p&gt;

&lt;p&gt;When I push changes to the repo, they don’t need to reinstall anything. One command syncs everything the plugin brings in, agents, skills, MCP config, all of it:&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/agentplugins/update-plugins-command.png&quot; alt=&quot;Running Chat: Update Plugins to pull the latest changes&quot; /&gt;&lt;/p&gt;

&lt;p&gt;That’s it. That’s plug-and-play, and nobody had to build a custom extension to get there.&lt;/p&gt;

&lt;p&gt;If you’re on &lt;em&gt;Claude Code&lt;/em&gt; instead of VS Code, same idea, just a terminal UI instead of a panel. &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;/plugin&lt;/code&gt; gets you the same tabs: &lt;em&gt;Marketplaces&lt;/em&gt; to register the repo, &lt;em&gt;Discover&lt;/em&gt; to find and install a plugin, &lt;em&gt;Installed&lt;/em&gt; to see what you’ve got and check for updates:&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/agentplugins/cc-plugin-marketplaces-tab.png&quot; alt=&quot;Adding a marketplace from the Claude Code /plugin UI&quot; /&gt;&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/agentplugins/cc-plugin-discover-tab.png&quot; alt=&quot;Discovering hello-plugin in the marketplace&quot; /&gt;&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/agentplugins/cc-plugin-installed-tab.png&quot; alt=&quot;hello-plugin installed, with its agents, skills, and MCP servers all listed&quot; /&gt;&lt;/p&gt;

&lt;hr /&gt;

&lt;h3 id=&quot;whats-actually-in-the-repository&quot;&gt;What’s actually in the repository&lt;/h3&gt;

&lt;p&gt;I put together a stripped-down version of what I now use as a template, &lt;strong&gt;&lt;em&gt;&lt;a href=&quot;https://github.com/tinestaric/plugin-marketplace-sample&quot;&gt;plugin-marketplace-sample&lt;/a&gt;&lt;/em&gt;&lt;/strong&gt;, so you can see the whole thing at once instead of just screenshots. Two kinds of file make this work: &lt;strong&gt;marketplace manifests&lt;/strong&gt; and &lt;strong&gt;plugin manifests&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;At the repo root, you need a marketplace manifest per ecosystem, since each one looks in a different spot:&lt;/p&gt;

&lt;div class=&quot;language-plaintext highlighter-rouge&quot;&gt;&lt;div class=&quot;highlight&quot;&gt;&lt;pre class=&quot;highlight&quot;&gt;&lt;code&gt;.github/plugin/marketplace.json      (GitHub Copilot)
.claude-plugin/marketplace.json      (Claude Code)
&lt;/code&gt;&lt;/pre&gt;&lt;/div&gt;&lt;/div&gt;

&lt;p&gt;Both files say the same thing, just to a different audience. A name for the marketplace, an owner, and a list of the plugins it serves, each one pointing at the folder it lives in:&lt;/p&gt;

&lt;div class=&quot;language-json highlighter-rouge&quot;&gt;&lt;div class=&quot;highlight&quot;&gt;&lt;pre class=&quot;highlight&quot;&gt;&lt;code&gt;&lt;span class=&quot;p&quot;&gt;{&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
  &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;name&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;s2&quot;&gt;&quot;sample-marketplace&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
  &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;owner&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;p&quot;&gt;{&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;name&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;s2&quot;&gt;&quot;Your Org&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;url&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;s2&quot;&gt;&quot;https://github.com/your-org&quot;&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;p&quot;&gt;},&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
  &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;plugins&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;p&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
    &lt;/span&gt;&lt;span class=&quot;p&quot;&gt;{&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
      &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;name&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;s2&quot;&gt;&quot;hello-plugin&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
      &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;source&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;s2&quot;&gt;&quot;./plugins/hello-plugin&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
      &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;description&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;s2&quot;&gt;&quot;Minimal example plugin demonstrating a skill and an agent shared between Claude Code and GitHub Copilot&quot;&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
    &lt;/span&gt;&lt;span class=&quot;p&quot;&gt;}&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
  &lt;/span&gt;&lt;span class=&quot;p&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;}&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/div&gt;&lt;/div&gt;

&lt;p&gt;That’s the whole catalog. This is the file that turns a plain repo into something you can register as a marketplace in &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;settings.json&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;The actual plugin lives in the folder that &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;source&lt;/code&gt; points to, and it carries its own manifest, again one per ecosystem:&lt;/p&gt;

&lt;div class=&quot;language-plaintext highlighter-rouge&quot;&gt;&lt;div class=&quot;highlight&quot;&gt;&lt;pre class=&quot;highlight&quot;&gt;&lt;code&gt;plugins/hello-plugin/
  plugin.json                  (Copilot plugin manifest)
  .claude-plugin/
    plugin.json                (Claude Code plugin manifest)
  agents/                      (shared by both tools)
    hello-agent.agent.md
  skills/                      (shared by both tools)
    hello-skill/
      SKILL.md
  .mcp.json                    (MCP servers, shared)
&lt;/code&gt;&lt;/pre&gt;&lt;/div&gt;&lt;/div&gt;

&lt;p&gt;One &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;agents/&lt;/code&gt; folder, one &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;skills/&lt;/code&gt; folder, nothing per-tool. Both &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;plugin.json&lt;/code&gt; files stay this minimal, no path overrides, because &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;agents/&lt;/code&gt; and &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;skills/&lt;/code&gt; are already the default folder names both tools look for:&lt;/p&gt;

&lt;div class=&quot;language-json highlighter-rouge&quot;&gt;&lt;div class=&quot;highlight&quot;&gt;&lt;pre class=&quot;highlight&quot;&gt;&lt;code&gt;&lt;span class=&quot;p&quot;&gt;{&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
  &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;name&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;s2&quot;&gt;&quot;hello-plugin&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
  &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;description&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;s2&quot;&gt;&quot;Example plugin: one skill, one agent, shared between Copilot and Claude Code&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
  &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;version&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;s2&quot;&gt;&quot;0.1.0&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
  &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;author&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;p&quot;&gt;{&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;nl&quot;&gt;&quot;name&quot;&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;s2&quot;&gt;&quot;Your Org&quot;&lt;/span&gt;&lt;span class=&quot;w&quot;&gt; &lt;/span&gt;&lt;span class=&quot;p&quot;&gt;}&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
&lt;/span&gt;&lt;span class=&quot;p&quot;&gt;}&lt;/span&gt;&lt;span class=&quot;w&quot;&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/div&gt;&lt;/div&gt;

&lt;p&gt;That’s genuinely all there is to the packaging itself: name, description, version, author. From there it’s exactly what you’d expect: agent &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;.md&lt;/code&gt; files with their usual frontmatter, skill folders with their &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;SKILL.md&lt;/code&gt;, an &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;.mcp.json&lt;/code&gt; if the plugin brings its own MCP servers. Nothing about authoring an individual agent or skill changes, you’re just putting them somewhere that can be distributed.&lt;/p&gt;

&lt;p&gt;This works over a plain &lt;em&gt;GitHub&lt;/em&gt; repo or an &lt;em&gt;Azure DevOps&lt;/em&gt; repo, whichever your org already uses for source control. No new infrastructure.&lt;/p&gt;

&lt;hr /&gt;

&lt;h3 id=&quot;why-theres-only-one-agents-folder&quot;&gt;Why there’s only one &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;agents/&lt;/code&gt; folder&lt;/h3&gt;

&lt;p&gt;&lt;em&gt;GitHub Copilot&lt;/em&gt; and &lt;em&gt;Claude Code&lt;/em&gt; name their built-in tools differently, Copilot’s &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;edit&lt;/code&gt; is Claude’s &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;Edit&lt;/code&gt; plus &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;Write&lt;/code&gt;. My first instinct was to work around that the obvious way: one agent folder per tool, &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;agents-copilot/&lt;/code&gt; and &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;agents-claude/&lt;/code&gt;, each with its own tool names in the frontmatter, each listed in its own &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;plugin.json&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;That broke the moment I actually tested it in VS Code. Copilot’s CLI only reads the &lt;em&gt;first&lt;/em&gt; &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;plugin.json&lt;/code&gt; it finds for a plugin, so the split worked fine there. But the Copilot extension inside VS Code reads &lt;em&gt;every&lt;/em&gt; &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;plugin.json&lt;/code&gt; it finds and merges them, so it picked up both folders at once and installed every agent twice.&lt;/p&gt;

&lt;p&gt;The fix is simpler than the workaround, and you don’t have to give up restricting tools to get it. List both platforms’ tool names together in one &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;tools&lt;/code&gt; array:&lt;/p&gt;

&lt;div class=&quot;language-yaml highlighter-rouge&quot;&gt;&lt;div class=&quot;highlight&quot;&gt;&lt;pre class=&quot;highlight&quot;&gt;&lt;code&gt;&lt;span class=&quot;na&quot;&gt;tools&lt;/span&gt;&lt;span class=&quot;pi&quot;&gt;:&lt;/span&gt; &lt;span class=&quot;pi&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;nv&quot;&gt;read&lt;/span&gt;&lt;span class=&quot;pi&quot;&gt;,&lt;/span&gt; &lt;span class=&quot;nv&quot;&gt;search&lt;/span&gt;&lt;span class=&quot;pi&quot;&gt;,&lt;/span&gt; &lt;span class=&quot;nv&quot;&gt;Read&lt;/span&gt;&lt;span class=&quot;pi&quot;&gt;,&lt;/span&gt; &lt;span class=&quot;nv&quot;&gt;Glob&lt;/span&gt;&lt;span class=&quot;pi&quot;&gt;]&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;&lt;/div&gt;&lt;/div&gt;

&lt;p&gt;Each platform only recognizes its own names and ignores the rest, so Copilot resolves &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;read&lt;/code&gt;/&lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;search&lt;/code&gt; and Claude Code resolves &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;Read&lt;/code&gt;/&lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;Glob&lt;/code&gt;, from the exact same file. One &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;agents/&lt;/code&gt; folder, one file per agent, read as-is by both &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;plugin.json&lt;/code&gt; manifests through their shared default.&lt;/p&gt;

&lt;p&gt;One constraint worth knowing before you set this up yourself: &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;skills/&lt;/code&gt; needs a subfolder per skill (&lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;skills/hello-skill/SKILL.md&lt;/code&gt;, not a loose &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;hello-skill.md&lt;/code&gt;), and inside that folder you can nest whatever you want, scripts, references, more subfolders, no problem. What you can’t do is nest the skill folders themselves any deeper, &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;skills/hello-skill/&lt;/code&gt; works, &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;skills/category/hello-skill/&lt;/code&gt; doesn’t. Ask me how I found that out.&lt;/p&gt;

&lt;hr /&gt;

&lt;h3 id=&quot;where-this-leaves-things&quot;&gt;Where this leaves things&lt;/h3&gt;

&lt;p&gt;Plugins don’t tell you what should be inside the box, that’s still on you and your team to figure out. But they remove the excuse for not shipping a default set at all. You’re no longer choosing between “build everyone their own extension” and “make people fend for themselves.” A plugin repo costs you a couple of manifest files and some discipline about where things live, and from there it’s one install command for anyone on your team, in whichever editor they already use.&lt;/p&gt;

&lt;p&gt;Go build that one agent from vol.3. Then put it somewhere the rest of your team can actually reach. Clone &lt;strong&gt;&lt;em&gt;&lt;a href=&quot;https://github.com/tinestaric/plugin-marketplace-sample&quot;&gt;the sample repo&lt;/a&gt;&lt;/em&gt;&lt;/strong&gt; if you want a starting point.&lt;/p&gt;

</description>
        <pubDate>Fri, 03 Jul 2026 10:00:00 +0200</pubDate>
        <link>https://tine.staric.net/blog/2026/agent-plugins/</link>
        <guid isPermaLink="true">https://tine.staric.net/blog/2026/agent-plugins/</guid>
        
        
        <category>al</category>
        
      </item>
    
      <item>
        <title>Copilot is just a fancy autocomplete (vol.3) - Steps towards agentic development</title>
        <description>&lt;p&gt;&lt;strong&gt;&lt;em&gt;&lt;a href=&quot;/blog/2025/copilot-fancy-autocomplete/&quot;&gt;Part 1&lt;/a&gt;&lt;/em&gt;&lt;/strong&gt; was about autocomplete, Edits, and Inline Chat. &lt;strong&gt;&lt;em&gt;&lt;a href=&quot;/blog/2025/copilot-fancy-autocomplete-vol2/&quot;&gt;Part 2&lt;/a&gt;&lt;/em&gt;&lt;/strong&gt; moved on to Agent mode and MCP servers, and closed with two promises: I’d talk about &lt;strong&gt;background agents&lt;/strong&gt; next, and a warning not to go overboard installing every MCP you can find, because it eats your context window.&lt;/p&gt;

&lt;p&gt;This post is really an extension of both, but the direct spark for it was this year’s &lt;strong&gt;&lt;em&gt;“Copilot is just a fancy autocomplete vol.2”&lt;/em&gt;&lt;/strong&gt; session, the one I’ve been touring conferences with. Somewhere between the first stop and the last, most of what I was presenting stopped being true, and that’s basically the whole premise of what follows.&lt;/p&gt;

&lt;p&gt;Both of those threads from vol.2, background agents and the MCP warning, are what this post is about. But first, a small detour.&lt;/p&gt;

&lt;p&gt;You’ve probably seen the LinkedIn posts by now. Someone shares an idea in the morning, and by lunch they’ve got tests, code, and documentation ready to review. &lt;a href=&quot;https://www.linkedin.com/in/vjeko/&quot;&gt;Vjeko&lt;/a&gt;, one of the sharpest people in the agentic development space right now, has a workflow where he goes from a feature idea to a fully tested, documented implementation in about &lt;strong&gt;24 minutes&lt;/strong&gt;. Not 24 minutes of watching a progress bar either: tests written first, code written to pass them, mutation testing to make sure the tests actually test something, docs generated at the end.&lt;/p&gt;

&lt;p&gt;That’s real. But it also feels a bit too mythical, and so we stay exactly where we are, one prompt at a time in the chat box on the right.&lt;/p&gt;

&lt;p&gt;I’m not trying to get you to that 24-minute workflow by tomorrow. I just want to show you the steps that get you closer, one at a time, from wherever you’re standing right now.&lt;/p&gt;

&lt;hr /&gt;

&lt;h3 id=&quot;everything-from-vol2-is-already-obsolete&quot;&gt;Everything from vol.2 is already obsolete&lt;/h3&gt;

&lt;p&gt;If I show you Visual Studio Code right now, it’s this:&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol3/agent-plan-modes.png&quot; alt=&quot;Agent and Plan modes, no Edit mode&quot; /&gt;&lt;/p&gt;

&lt;p&gt;No Edit mode anymore. Just gone. The feature I called the &lt;em&gt;crown jewel&lt;/em&gt; twelve months ago, the chat box where you type requirements and it generates code, quietly removed. If that doesn’t tell you how fast this moves, I don’t know what will.&lt;/p&gt;

&lt;p&gt;Two things did survive the last twelve months though:&lt;/p&gt;

&lt;ul&gt;
  &lt;li&gt;Claude models are still the best. Claude Sonnet and Opus, they’re the best for AL. Though, I’ve seen Gemini 3.1 Pro come very close.&lt;/li&gt;
  &lt;li&gt;Taking control of your context still matters. Agents are decent at finding their own files now, but pointing them at exactly the right file still gets you better results than letting them guess.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Everything else from vol.2, I’d rewrite.&lt;/p&gt;

&lt;hr /&gt;

&lt;h3 id=&quot;the-mcp-overload-problem-revisited&quot;&gt;The MCP overload problem, revisited&lt;/h3&gt;

&lt;p&gt;In vol.2 I explained MCP servers with the USB-C analogy: tool boxes you hand your agent so it can reach past your workspace, into your ticket tracker, your docs, your telemetry. Go read that post first if you haven’t, it’s still the best explanation I’ve got.&lt;/p&gt;

&lt;p&gt;Quick reminder of why they matter at all, though. Ask an agent with no tools attached about a work item:&lt;/p&gt;

&lt;blockquote&gt;
  &lt;p&gt;What is task 107308 about?&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol3/no-mcp-no-clue.png&quot; alt=&quot;An agent with no MCP has no idea what task 107308 is&quot; /&gt;&lt;/p&gt;

&lt;p&gt;No idea. It doesn’t even know where to look. Give it the Azure DevOps MCP:&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol3/enable-devops-mcp.png&quot; alt=&quot;Enabling the Azure DevOps MCP for the agent&quot; /&gt;&lt;/p&gt;

&lt;p&gt;Ask again, and now it goes and actually fetches the ticket:&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol3/devops-mcp-result.png&quot; alt=&quot;The same agent now finds and summarizes task 107308 via the Azure DevOps MCP&quot; /&gt;&lt;/p&gt;

&lt;p&gt;That’s the whole point of MCPs in one before-and-after.&lt;/p&gt;

&lt;p&gt;Now, the community kept building since vol.2, and the full, up to date list is on the &lt;a href=&quot;https://alguidelines.dev/docs/agentic-coding/communityresources/tools/&quot;&gt;AL Guidelines agentic tools page&lt;/a&gt;. One I’d specifically call out: &lt;strong&gt;TelemetryBuddy&lt;/strong&gt;, by &lt;a href=&quot;https://www.linkedin.com/in/ericwauters&quot;&gt;waldo&lt;/a&gt;, turns &lt;em&gt;“what are the top 10 customers with errors in job queue entries”&lt;/em&gt; into a KQL query against Application Insights, runs it, and hands you back a readable answer. No KQL knowledge required anymore.&lt;/p&gt;

&lt;p&gt;Now, let’s pick up exactly where vol.2 left off. I told you not to install everything. Let’s actually see why.&lt;/p&gt;

&lt;p&gt;I gave my agent every MCP I had, including one for Python even though I was sitting in an AL workspace.&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol3/all-tools-enabled.png&quot; alt=&quot;Every single tool enabled at once&quot; /&gt;&lt;/p&gt;

&lt;p&gt;Then I typed the most demanding prompt I could think of:&lt;/p&gt;

&lt;blockquote&gt;
  &lt;p&gt;Hello&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol3/hello-prompt.png&quot; alt=&quot;Hello prompt to an overloaded agent&quot; /&gt;&lt;/p&gt;

&lt;p&gt;Go on, guess how many tokens that cost.&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol3/hello-tokens.png&quot; alt=&quot;24,000 tokens spent on saying hello&quot; /&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;24,000 tokens.&lt;/strong&gt; For “hello.” Sent straight to the model with no tools attached, it would’ve cost about 10.&lt;/p&gt;

&lt;p&gt;Now, Sonnet’s default context window recently jumped to 1M tokens, so 24K doesn’t threaten to fill it up the way it would’ve a few months ago. But that doesn’t mean it’s free. First, it’s still 24,000 tokens of tool definitions you’re paying for and waiting on, for every single message, whether or not the agent ever touches half those tools. Second, models don’t actually reason better with more junk sitting in front of them. Stuff a context window with tool definitions it doesn’t need, and the model has to work around that noise, even if it technically still fits. A window that isn’t full isn’t the same thing as a window that’s clean.&lt;/p&gt;

&lt;p&gt;And the old danger hasn’t disappeared, it’s just further away now. Fill that window up regardless, and once you hit the last ~10% left, the agent has to start &lt;em&gt;summarizing chat history&lt;/em&gt; to make room. Summarizing means throwing information out. An agent that knew exactly how you like your code written suddenly feels lobotomized. Wrong file names, wrong properties, mistakes it wasn’t making five minutes earlier.&lt;/p&gt;

&lt;p&gt;An agent without any tools is going to perform basically the same way ChatGPT would if we just jammed it in Visual Studio Code, not really useful. And now you’ve seen an agent with every tool also fail, just in the other direction. So what’s the answer?&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Custom agents.&lt;/strong&gt;&lt;/p&gt;

&lt;hr /&gt;

&lt;h3 id=&quot;step-1-custom-agents&quot;&gt;Step 1: Custom agents&lt;/h3&gt;

&lt;p&gt;If last year’s crown jewel was Edit mode, this year’s is custom agents. Custom agents are the first step you take closer to agentic development, because building one forces you to think about your workflow. What are you actually doing, day to day? What does your process look like? You break it down into individual parts, and you create an agent for each part.&lt;/p&gt;

&lt;ul&gt;
  &lt;li&gt;Planning&lt;/li&gt;
  &lt;li&gt;Writing code&lt;/li&gt;
  &lt;li&gt;Writing tests&lt;/li&gt;
  &lt;li&gt;Reviewing a backlog item&lt;/li&gt;
  &lt;li&gt;Writing docs&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Each of those can be its own agent, its own persona, its own scoped-down set of tools. No telemetry MCP on your test-plan-helper. No admin MCP on your backlog reviewer. They don’t need it.&lt;/p&gt;

&lt;p&gt;Used to be you’d start from a blank template and fill in the prompt and tool list by hand. There’s a better way now. GitHub Copilot has a &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;/create-agent&lt;/code&gt; skill: describe the job in plain language and let an agent build the agent for you.&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol3/create-agent-prompt.png&quot; alt=&quot;Creating a custom agent with a plain-language description&quot; /&gt;&lt;/p&gt;

&lt;p&gt;A minute later, a working backlog reviewer: checking requirement quality, acceptance criteria, flagging risks and dependencies.&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol3/backlog-reviewer-agent.png&quot; alt=&quot;Generated backlog reviewer agent definition&quot; /&gt;&lt;/p&gt;

&lt;p&gt;Won’t be perfect on the first try. It’ll be the perfect first step. From here you iterate, add a tool, tighten the persona, and eventually my prompt for reviewing a backlog item is three words.&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol3/three-word-review.png&quot; alt=&quot;Three-word prompt reviewing a task in full&quot; /&gt;&lt;/p&gt;

&lt;p&gt;Everything about what the agent should do and how it should use its tools already lives in the agent definition, I’m not re-explaining it every time. Also solves onboarding, by the way. Hand a colleague a super-agent that does five things poorly, they won’t come back to it. Hand them one agent that does backlog review well, and they’ll use it every time they need backlog review, then reach for a different one for the next job.&lt;/p&gt;

&lt;hr /&gt;

&lt;h3 id=&quot;step-2-skills&quot;&gt;Step 2: Skills&lt;/h3&gt;

&lt;p&gt;An agent knows how to do some things. Not everything, and it shouldn’t have to carry all of it around at once.&lt;/p&gt;

&lt;p&gt;Take compiling and publishing an app, running tests, reporting back results. An agent today is smart enough to figure out that whole sequence itself, find the BC Container Helper, write the right PowerShell, run it. But it’ll burn a lot of tokens rediscovering that path every single time.&lt;/p&gt;

&lt;p&gt;That’s what skills solve, through something called &lt;em&gt;progressive disclosure&lt;/em&gt;. By default your agent only ever sees a skill’s name and a one-line description, barely any context cost at all. Only when a prompt actually needs it does it load the full instructions.&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol3/skill-example.png&quot; alt=&quot;Example skill file for compiling, publishing, and running tests&quot; /&gt;&lt;/p&gt;

&lt;p&gt;Rule of thumb: if a task takes your agent a handful of turns and a few minutes to figure out, ask it at the end to package that into a skill. Same idea as &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;/create-agent&lt;/code&gt;, there’s a &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;/create-skill&lt;/code&gt; command for this too, so you don’t have to hand-write the skill file yourself. Next time it won’t spend thousands of tokens rediscovering how to publish a workload, it’ll just know.&lt;/p&gt;

&lt;p&gt;This also kills something I mentioned in vol.2, sort of implicitly. Those old &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;.prompt.md&lt;/code&gt; stored-prompt files? I don’t use them anymore. Anything that used to be a pre-written prompt is now a skill, same &lt;em&gt;“run this as a slash command”&lt;/em&gt; behavior, except an agent can also discover and load it on its own when it’s relevant. A stored prompt never could do that.&lt;/p&gt;

&lt;p&gt;One distinction worth keeping straight. Your agent’s instructions (or your general &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;copilot-instructions.md&lt;/code&gt; or &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;CLAUDE.md&lt;/code&gt;) hold what’s relevant &lt;em&gt;every single time&lt;/em&gt;: architecture, where things live, how you like code written. Skills hold what’s relevant &lt;em&gt;sometimes&lt;/em&gt;, a how-to that only matters when the task calls for it. Don’t copy-paste your architecture notes into every agent’s skill file. That belongs in instructions, once.&lt;/p&gt;

&lt;hr /&gt;

&lt;h3 id=&quot;step-3-the-reflection-loop&quot;&gt;Step 3: The reflection loop&lt;/h3&gt;

&lt;p&gt;This is, in my opinion, the most important habit in this whole post.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Fix the agent. Don’t fix the code.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol3/fix-the-agent.png&quot; alt=&quot;Don&apos;t fix yourself. Fix the agent.&quot; /&gt;&lt;/p&gt;

&lt;p&gt;Old failure mode: agent mode makes a mistake, you sigh, take over, fix it manually. You get a little better at spotting that mistake. The agent stays exactly as dumb as before, and it’ll make the same mistake next week.&lt;/p&gt;

&lt;p&gt;Here’s the loop instead. Over one longer session, the agent kept making small mistakes:&lt;/p&gt;

&lt;ul&gt;
  &lt;li&gt;Auto-increment on a temporary table&lt;/li&gt;
  &lt;li&gt;A tooltip that belonged on the table ending up on the page instead&lt;/li&gt;
  &lt;li&gt;A property that should’ve lived on the header sitting on a field&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Each time, I didn’t touch the code myself, I told the agent what was wrong and had it fix it. But only at the very end of the conversation did I fire off the actual reflection loop prompt:&lt;/p&gt;

&lt;blockquote&gt;
  &lt;p&gt;Based on our conversation and the fixes I made you make, what would you fix in your instructions and skills so the next time you don’t make the same mistake?&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;It goes back through the conversation, identifies what could’ve gone better, and proposes fixes. &lt;em&gt;Auto-increment should never be used on temporary tables, put that in instructions.&lt;/em&gt; &lt;em&gt;Tooltips belong on the field, not the page, double check that going forward.&lt;/em&gt; &lt;em&gt;Same for properties that belong on the header, not the field.&lt;/em&gt; Doesn’t really matter what specifically comes out of it. What matters is the habit: instruct the agent how to fix the mistake, let it fix the skill or instructions itself, then move on.&lt;/p&gt;

&lt;p&gt;Do that consistently, and in a few months you’ve got agents that write code the way you would.&lt;/p&gt;

&lt;hr /&gt;

&lt;h3 id=&quot;step-4-sub-agents&quot;&gt;Step 4: Sub-agents&lt;/h3&gt;

&lt;p&gt;Stumbled into this one about a week after building my first backlog reviewer. A consultant gave it a real task: build a transfer order approval workflow, similar to the existing sales order one.&lt;/p&gt;

&lt;p&gt;Sounded simple. I told the agent to load the base app code for sales order approval via the &lt;a href=&quot;https://github.com/StefanMaron/AL-Dependency-MCP-Server&quot;&gt;AL Symbols MCP&lt;/a&gt;, pull the official documentation, and bring in a specialist from &lt;a href=&quot;https://github.com/JeremyVyska/bc-code-intelligence-mcp&quot;&gt;BC Code Intelligence&lt;/a&gt; before we even started on the design. Except it never got that far. The moment all three loaded, the context window filled up and it summarized the chat history before I could say a word about the actual design.&lt;/p&gt;

&lt;p&gt;That’s when sub-agents clicked. The main agent can spin up new agents with a completely empty context window of their own. One digs through the base app. One goes through the docs. One consults the specialist. Each starts fresh with the full window available, and when it’s done, it doesn’t hand back everything it read, it hands back a condensed research report. Which files matter, which procedures matter, how they work.&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol3/sub-agents-parallel.png&quot; alt=&quot;Sub-agents analyzing base app, docs, and a specialist in parallel&quot; /&gt;&lt;/p&gt;

&lt;p&gt;Here’s what that actually looked like. First, the agent itself needs the tools for this, &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;agent&lt;/code&gt; and &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;runSubagent&lt;/code&gt;, switched on. Nothing spins up in parallel without them:&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol3/subagent-tools-enabled.png&quot; alt=&quot;Enabling the agent and runSubagent tools on a custom agent&quot; /&gt;&lt;/p&gt;

&lt;p&gt;With those enabled, same prompt as before, load the docs, check the base app via the AL Symbols MCP, ask a specialist, except this time each of those runs as its own sub-agent instead of stacking up in my main conversation:&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol3/sub-agents-prompt.png&quot; alt=&quot;Prompt spinning up three sub-agents in parallel&quot; /&gt;&lt;/p&gt;

&lt;p&gt;Zoom into just one of them, the one digging through the base app, and you can see it doing exactly what I’d have done myself: pulling packages, searching objects, reading through the workflow code, one tool call after another.&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol3/sub-agent-tool-calls.png&quot; alt=&quot;One sub-agent&apos;s tool calls while researching the base app&quot; /&gt;&lt;/p&gt;

&lt;p&gt;None of that noise comes back to my main agent though. What comes back is this:&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol3/sub-agent-research-report.png&quot; alt=&quot;The condensed research report handed back to the main agent&quot; /&gt;&lt;/p&gt;

&lt;p&gt;A research report, not a transcript. Object IDs, source files, key procedures, exactly the shape of information I actually needed to start talking about the transfer order workflow.&lt;/p&gt;

&lt;p&gt;Parallelism is a nice bonus, three things happening at once instead of one after another. But that’s not the actual win. The win is my main agent’s context window barely moved. All it got was a tight summary, and there was still plenty of room left to actually talk through how to implement the workflow.&lt;/p&gt;

&lt;p&gt;Once you’re here, the next realization comes on its own: your custom agents don’t just call sub-agents for research, they can &lt;em&gt;be&lt;/em&gt; sub-agents themselves. Planner calls code writer, code writer calls test writer, test writer calls documentation writer. Congratulations, you’ve built an orchestrator, and this is more or less what gets you to a Vjeko-style, idea-to-production workflow.&lt;/p&gt;

&lt;hr /&gt;

&lt;h3 id=&quot;step-5-background-agents&quot;&gt;Step 5: Background agents&lt;/h3&gt;

&lt;p&gt;Two things to untangle before this one makes sense.&lt;/p&gt;

&lt;p&gt;Claude Sonnet is a model, not an agent. GitHub Copilot in VS Code is an &lt;em&gt;agent framework&lt;/em&gt; (agent harness, officially, but framework fits my head better) that builds agents on top of that model. Cursor is a different framework, Claude Code is another, and each behaves differently even pointed at the exact same model, because the framework shapes the behavior. Copilot CLI is yet another one of these, it just lives in your terminal instead of your editor, and because it lives there, you can spin up as many as you want.&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol3/agent-frameworks.png&quot; alt=&quot;Task, model, and tools, the same shape across VS Code, GitHub, CLI, and other agent frameworks&quot; /&gt;&lt;/p&gt;

&lt;p&gt;Second thing, once you have multiple agents running, you hit the same problem developers had before branching existed. Two agents on the same files start stepping on each other’s toes, fast. The fix is the same one Git gave us, repurposed: &lt;strong&gt;git worktrees&lt;/strong&gt;. Think of a worktree as a branch that’s also a folder. When an agent branches out, it copies the files into its own folder and works there undisturbed, and only brings the results back when done. The agent equivalent of a pull request.&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol3/git-worktrees.png&quot; alt=&quot;Worktrees as folders sitting alongside branches in the same repo&quot; /&gt;&lt;/p&gt;

&lt;p&gt;With both in place, here’s what a background task looks like. I had 600 files with tooltips like &lt;em&gt;“specifies the value of field X.”&lt;/em&gt; Nobody’s ever going to pay you to write better tooltips, but it’s the perfect task for an agent. Do I really want to watch an agent go through 600 files and adjust all of them though? I don’t.&lt;/p&gt;

&lt;p&gt;Right next to where you type your prompt, there’s a target selector, &lt;em&gt;Local&lt;/em&gt; by default:&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol3/target-local.png&quot; alt=&quot;Target selector set to Local by default&quot; /&gt;&lt;/p&gt;

&lt;p&gt;Switch it to &lt;em&gt;Copilot CLI&lt;/em&gt;:&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol3/target-dropdown.png&quot; alt=&quot;Switching the target to Copilot CLI&quot; /&gt;&lt;/p&gt;

&lt;p&gt;And once you’re on CLI, a second selector shows up for isolation mode. Switch that to &lt;em&gt;Worktree&lt;/em&gt;:&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol3/worktree-isolation-mode.png&quot; alt=&quot;Switching isolation mode to Worktree&quot; /&gt;&lt;/p&gt;

&lt;p&gt;So I used Copilot CLI, gave it the prompt, set the isolation mode to &lt;em&gt;worktree&lt;/em&gt;, let it go:&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol3/tooltip-prompt.png&quot; alt=&quot;Sending the tooltip prompt to Copilot CLI, isolated worktree created immediately&quot; /&gt;&lt;/p&gt;

&lt;p&gt;Took a while, didn’t care, I wasn’t watching it. When it finished, same diff view you’d expect, 96 files, +218/-218. Look at the left side though: my actual workspace, Source Control panel, empty. Nothing to sync, nothing to commit. Every one of those file paths on the right points into a separate worktree folder, not my project. It’s not in my workspace until I click Apply.&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol3/background-tooltips.png&quot; alt=&quot;96 files changed in an isolated worktree, source control in the actual workspace untouched&quot; /&gt;&lt;/p&gt;

&lt;p&gt;I want to be clear about what this actually was for me though. Running Copilot CLI from inside VS Code isn’t the end goal, it was the stepping stone that got me comfortable running agents directly in the terminal instead of watching a chat pane. Compared to just opening a bunch of GitHub Copilot sessions in VS Code, which you can already do, the real benefit of CLI is &lt;strong&gt;worktree isolation&lt;/strong&gt;. Multiple VS Code sessions still fight over the same files. CLI in worktree mode doesn’t.&lt;/p&gt;

&lt;p&gt;And once you’re comfortable with that, it’s worth just opening Copilot CLI on its own, no VS Code around it at all:&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol3/copilot-cli-terminal.png&quot; alt=&quot;GitHub Copilot CLI running directly in the terminal&quot; /&gt;&lt;/p&gt;

&lt;p&gt;This is genuinely how I work day to day now. A handful of terminal windows open, each running an agent on a completely unrelated task. Why would I sit and watch my orchestrator orchestrate, when the only thing I care about is reviewing the final output?&lt;/p&gt;

&lt;p&gt;And it doesn’t have to stop at your own terminal. Because it’s a command, and a command is something you can put in a script. And a script can live in a pipeline. Here’s what that looks like, a step in a larger script that shells out to &lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;copilot&lt;/code&gt; for a code review pass:&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol3/copilot-cli-pipeline-script.png&quot; alt=&quot;A PowerShell pipeline step invoking Copilot CLI for a code review&quot; /&gt;&lt;/p&gt;

&lt;p&gt;That’s how you get from “I run agents in a terminal” to actually automating pieces of your workflow, generating documentation, drafting release notes, maybe even a first pass at code review, without anyone sitting there watching it happen at all.&lt;/p&gt;

&lt;hr /&gt;

&lt;h3 id=&quot;where-this-leaves-you&quot;&gt;Where this leaves you&lt;/h3&gt;

&lt;p&gt;Custom agents, skills, the reflection loop, sub-agents, background agents. Five steps, and you don’t have to sprint through all of them this month, or even this quarter. What matters is you don’t stop wherever you happen to land. Take the next step, then the one after that, and eventually you’re a lot closer to that 24-minute workflow than you think.&lt;/p&gt;

&lt;p&gt;I know I’m leaving some things dangling, and I’m not going to pretend I’ve covered everything here either. There’s going to be a &lt;strong&gt;vol.4&lt;/strong&gt;. There’s basically always going to be a next volume, at the pace this moves. Honestly, by the time you’re reading this, some of what’s above has probably already aged like milk.&lt;/p&gt;

&lt;p&gt;Until then, go build yourself one agent. Just one. It doesn’t need to be the whole suite.&lt;/p&gt;

</description>
        <pubDate>Wed, 01 Jul 2026 10:00:00 +0200</pubDate>
        <link>https://tine.staric.net/blog/2026/copilot-fancy-autocomplete-vol3/</link>
        <guid isPermaLink="true">https://tine.staric.net/blog/2026/copilot-fancy-autocomplete-vol3/</guid>
        
        
        <category>al</category>
        
      </item>
    
      <item>
        <title>Format Cheatsheet</title>
        <description>&lt;p&gt;This is a &lt;strong&gt;very quick cheatsheet&lt;/strong&gt; for all the different &lt;strong&gt;&lt;em&gt;formats&lt;/em&gt;&lt;/strong&gt; that you get out of &lt;strong&gt;AL&lt;/strong&gt; using the &lt;strong&gt;&lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;Format()&lt;/code&gt; function&lt;/strong&gt; with various &lt;strong&gt;&lt;em&gt;format options.&lt;/em&gt;&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;I’ve recently been trying to again find the format that returns the &lt;strong&gt;GUID value&lt;/strong&gt; &lt;strong&gt;&lt;em&gt;without the curly braces&lt;/em&gt;&lt;/strong&gt;, and as I didn’t want to test all the options again, I finally created this &lt;strong&gt;cheatsheet&lt;/strong&gt; for myself. Maybe it will be &lt;strong&gt;&lt;em&gt;useful for you&lt;/em&gt;&lt;/strong&gt; as well.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Format 8&lt;/em&gt; was omitted from this table as it produces no meaningful formatting for any data type.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Click&lt;/em&gt; on any cell to &lt;em&gt;expand&lt;/em&gt; and see the &lt;em&gt;full value&lt;/em&gt;.&lt;/p&gt;

&lt;blockquote&gt;
  &lt;p&gt;&lt;strong&gt;&lt;em&gt;Note:&lt;/em&gt;&lt;/strong&gt;&lt;br /&gt;
&lt;em&gt;Values marked with an &lt;strong&gt;asterisk (*)&lt;/strong&gt; are &lt;strong&gt;language dependent&lt;/strong&gt; and to make that obvious, I left them in Spanish.&lt;/em&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;style&gt;
.format-table-wrapper {
  overflow-x: auto;
  -webkit-overflow-scrolling: touch;
  margin: 1.5em 0;
}
.format-table {
  border-collapse: collapse;
  width: 100%;
  min-width: 1000px;
  font-size: 0.85em;
  font-family: &apos;Courier New&apos;, Courier, monospace;
}
.format-table th,
.format-table td {
  border-bottom: 1px solid #ddd;
  padding: 10px 8px;
  text-align: left;
  white-space: nowrap;
  max-width: 200px;
  overflow: hidden;
  text-overflow: ellipsis;
  transition: all 0.2s ease;
}
.format-table th {
  font-weight: 600;
  border-bottom: 2px solid #999;
  background: white;
  font-family: -apple-system, BlinkMacSystemFont, &apos;Segoe UI&apos;, sans-serif;
}
.format-table th:first-child,
.format-table td:first-child {
  position: sticky;
  left: 0;
  background: white;
  font-weight: 500;
  z-index: 1;
  box-shadow: 2px 0 4px rgba(0,0,0,0.05);
  font-family: -apple-system, BlinkMacSystemFont, &apos;Segoe UI&apos;, sans-serif;
  min-width: 90px;
  max-width: 90px;
}
.format-table th:first-child {
  z-index: 2;
}
.format-table td:not(:first-child) {
  cursor: pointer;
}
.format-table td.expanded {
  white-space: normal;
  word-break: break-all;
  max-width: none;
  background: #f0f7ff;
  position: relative;
}
.format-table td.expanded::before {
  content: &apos;✕&apos;;
  position: absolute;
  top: 4px;
  right: 4px;
  font-size: 0.7em;
  color: #666;
  font-family: Arial, sans-serif;
}
&lt;/style&gt;

&lt;script&gt;
document.addEventListener(&apos;DOMContentLoaded&apos;, function() {
  const cells = document.querySelectorAll(&apos;.format-table td:not(:first-child)&apos;);
  cells.forEach(cell =&gt; {
    cell.addEventListener(&apos;click&apos;, function() {
      // Close all other expanded cells first
      cells.forEach(c =&gt; {
        if (c !== cell) c.classList.remove(&apos;expanded&apos;);
      });
      // Toggle current cell
      this.classList.toggle(&apos;expanded&apos;);
    });
  });
});
&lt;/script&gt;

&lt;div class=&quot;format-table-wrapper&quot;&gt;
  &lt;table class=&quot;format-table&quot;&gt;
    &lt;thead&gt;
      &lt;tr&gt;
        &lt;th&gt;Data Type&lt;/th&gt;
        &lt;th&gt;Format 0&lt;/th&gt;
        &lt;th&gt;Format 1&lt;/th&gt;
        &lt;th&gt;Format 2&lt;/th&gt;
        &lt;th&gt;Format 3&lt;/th&gt;
        &lt;th&gt;Format 4&lt;/th&gt;
        &lt;th&gt;Format 5&lt;/th&gt;
        &lt;th&gt;Format 6&lt;/th&gt;
        &lt;th&gt;Format 7&lt;/th&gt;
        &lt;th&gt;Format 9&lt;/th&gt;
      &lt;/tr&gt;
    &lt;/thead&gt;
    &lt;tbody&gt;
      &lt;tr&gt;
        &lt;td&gt;DateTime&lt;/td&gt;
        &lt;td&gt;10/07/25 02:30 PM&lt;/td&gt;
        &lt;td&gt;10/07/25 02:30:52 PM&lt;/td&gt;
        &lt;td&gt;10/07/25 02:30 PM&lt;/td&gt;
        &lt;td&gt;10/07/25 02:30:52 PM&lt;/td&gt;
        &lt;td&gt;N/A&lt;/td&gt;
        &lt;td&gt;N/A&lt;/td&gt;
        &lt;td&gt;N/A&lt;/td&gt;
        &lt;td&gt;N/A&lt;/td&gt;
        &lt;td&gt;2025-10-07T11:30:52Z&lt;/td&gt;
      &lt;/tr&gt;
      &lt;tr&gt;
        &lt;td&gt;Date&lt;/td&gt;
        &lt;td&gt;10/07/2025&lt;/td&gt;
        &lt;td&gt;10/07/2025&lt;/td&gt;
        &lt;td&gt;100725D&lt;/td&gt;
        &lt;td&gt;25/10/07&lt;/td&gt;
        &lt;td&gt;*Octubre 7, 2025&lt;/td&gt;
        &lt;td&gt;100725&lt;/td&gt;
        &lt;td&gt;251007&lt;/td&gt;
        &lt;td&gt;*Oct 7, 2025&lt;/td&gt;
        &lt;td&gt;07/10/2025&lt;/td&gt;
      &lt;/tr&gt;
      &lt;tr&gt;
        &lt;td&gt;Time&lt;/td&gt;
        &lt;td&gt;2:30:52 PM&lt;/td&gt;
        &lt;td&gt;2:30:52 PM&lt;/td&gt;
        &lt;td&gt;143052T&lt;/td&gt;
        &lt;td&gt;N/A&lt;/td&gt;
        &lt;td&gt;N/A&lt;/td&gt;
        &lt;td&gt;N/A&lt;/td&gt;
        &lt;td&gt;N/A&lt;/td&gt;
        &lt;td&gt;N/A&lt;/td&gt;
        &lt;td&gt;14:30:52&lt;/td&gt;
      &lt;/tr&gt;
      &lt;tr&gt;
        &lt;td&gt;Guid&lt;/td&gt;
        &lt;td&gt;{FDBEA7C4-9205-4A3A-998E-D7B020529BC1}&lt;/td&gt;
        &lt;td&gt;{FDBEA7C4-9205-4A3A-998E-D7B020529BC1}&lt;/td&gt;
        &lt;td&gt;{FDBEA7C4-9205-4A3A-998E-D7B020529BC1}&lt;/td&gt;
        &lt;td&gt;FDBEA7C492054A3A998ED7B020529BC1&lt;/td&gt;
        &lt;td&gt;FDBEA7C4-9205-4A3A-998E-D7B020529BC1&lt;/td&gt;
        &lt;td&gt;(FDBEA7C4-9205-4A3A-998E-D7B020529BC1)&lt;/td&gt;
        &lt;td&gt;{0XFDBEA7C4,0X9205,0X4A3A,{0X99,0X8E,0XD7,0XB0,0X20,0X52,0X9B,0XC1}}&lt;/td&gt;
        &lt;td&gt;N/A&lt;/td&gt;
        &lt;td&gt;{FDBEA7C4-9205-4A3A-998E-D7B020529BC1}&lt;/td&gt;
      &lt;/tr&gt;
      &lt;tr&gt;
        &lt;td&gt;Decimal&lt;/td&gt;
        &lt;td&gt;1,234,567.89&lt;/td&gt;
        &lt;td&gt;1234567.89&lt;/td&gt;
        &lt;td&gt;1234567.89&lt;/td&gt;
        &lt;td&gt;1,234,567.89&lt;/td&gt;
        &lt;td&gt;1234567.89&lt;/td&gt;
        &lt;td&gt;N/A&lt;/td&gt;
        &lt;td&gt;N/A&lt;/td&gt;
        &lt;td&gt;N/A&lt;/td&gt;
        &lt;td&gt;1234567.89&lt;/td&gt;
      &lt;/tr&gt;
      &lt;tr&gt;
        &lt;td&gt;Boolean&lt;/td&gt;
        &lt;td&gt;*Sí&lt;/td&gt;
        &lt;td&gt;*Sí&lt;/td&gt;
        &lt;td&gt;1&lt;/td&gt;
        &lt;td&gt;N/A&lt;/td&gt;
        &lt;td&gt;N/A&lt;/td&gt;
        &lt;td&gt;N/A&lt;/td&gt;
        &lt;td&gt;N/A&lt;/td&gt;
        &lt;td&gt;N/A&lt;/td&gt;
        &lt;td&gt;TRUE&lt;/td&gt;
      &lt;/tr&gt;
      &lt;tr&gt;
        &lt;td&gt;Option&lt;/td&gt;
        &lt;td&gt;*Pedido abierto&lt;/td&gt;
        &lt;td&gt;*Pedido abierto&lt;/td&gt;
        &lt;td&gt;1&lt;/td&gt;
        &lt;td&gt;N/A&lt;/td&gt;
        &lt;td&gt;N/A&lt;/td&gt;
        &lt;td&gt;N/A&lt;/td&gt;
        &lt;td&gt;N/A&lt;/td&gt;
        &lt;td&gt;N/A&lt;/td&gt;
        &lt;td&gt;1&lt;/td&gt;
      &lt;/tr&gt;
      &lt;tr&gt;
        &lt;td&gt;Enum&lt;/td&gt;
        &lt;td&gt;*Pedido abierto&lt;/td&gt;
        &lt;td&gt;*Pedido abierto&lt;/td&gt;
        &lt;td&gt;1&lt;/td&gt;
        &lt;td&gt;N/A&lt;/td&gt;
        &lt;td&gt;N/A&lt;/td&gt;
        &lt;td&gt;N/A&lt;/td&gt;
        &lt;td&gt;N/A&lt;/td&gt;
        &lt;td&gt;N/A&lt;/td&gt;
        &lt;td&gt;1&lt;/td&gt;
      &lt;/tr&gt;
      &lt;tr&gt;
        &lt;td&gt;Duration&lt;/td&gt;
        &lt;td&gt;*1 hora 1 minuto 1 segundo&lt;/td&gt;
        &lt;td&gt;*1 hora 1 minuto 1 segundo&lt;/td&gt;
        &lt;td&gt;1 2 1 4 1 6&lt;/td&gt;
        &lt;td&gt;N/A&lt;/td&gt;
        &lt;td&gt;N/A&lt;/td&gt;
        &lt;td&gt;N/A&lt;/td&gt;
        &lt;td&gt;N/A&lt;/td&gt;
        &lt;td&gt;N/A&lt;/td&gt;
        &lt;td&gt;P0DT1H1M1.0S&lt;/td&gt;
      &lt;/tr&gt;
      &lt;tr&gt;
        &lt;td&gt;RecordId&lt;/td&gt;
        &lt;td&gt;Frmt. Tst. Tbl.: 1&lt;/td&gt;
        &lt;td&gt;Format Test Table: 1&lt;/td&gt;
        &lt;td&gt;Frmt. Tst. Tbl.: 1&lt;/td&gt;
        &lt;td&gt;N/A&lt;/td&gt;
        &lt;td&gt;N/A&lt;/td&gt;
        &lt;td&gt;N/A&lt;/td&gt;
        &lt;td&gt;N/A&lt;/td&gt;
        &lt;td&gt;N/A&lt;/td&gt;
        &lt;td&gt;[MyCompany.Tables.&quot;Frmt. Tst. Tbl.&quot;]: 1&lt;/td&gt;
      &lt;/tr&gt;
      &lt;tr&gt;
        &lt;td&gt;DateFormula&lt;/td&gt;
        &lt;td&gt;*1T+2D&lt;/td&gt;
        &lt;td&gt;*1T+2D&lt;/td&gt;
        &lt;td&gt;&amp;lt;1Q+2D&amp;gt;&lt;/td&gt;    
        &lt;td&gt;N/A&lt;/td&gt;
        &lt;td&gt;N/A&lt;/td&gt;
        &lt;td&gt;N/A&lt;/td&gt;
        &lt;td&gt;N/A&lt;/td&gt;
        &lt;td&gt;N/A&lt;/td&gt;
        &lt;td&gt;1Q+2D&lt;/td&gt;
      &lt;/tr&gt;
    &lt;/tbody&gt;
  &lt;/table&gt;
&lt;/div&gt;

</description>
        <pubDate>Tue, 07 Oct 2025 10:00:00 +0200</pubDate>
        <link>https://tine.staric.net/blog/2025/format-cheatsheet/</link>
        <guid isPermaLink="true">https://tine.staric.net/blog/2025/format-cheatsheet/</guid>
        
        
        <category>al</category>
        
      </item>
    
      <item>
        <title>Copilot is just a fancy autocomplete (vol.2)</title>
        <description>&lt;p&gt;A few weeks ago, I was asked if I’d have time to deliver the &lt;strong&gt;&lt;em&gt;“Copilot is just a fancy autocomplete…“&lt;/em&gt;&lt;/strong&gt; session from this year’s &lt;strong&gt;&lt;em&gt;&lt;a href=&quot;https://www.bctechdays.com/event&quot;&gt;BC TechDays&lt;/a&gt;&lt;/em&gt;&lt;/strong&gt; again for one of our partners at their internal event. As I was preparing, I realized that even in the span of &lt;strong&gt;3 months since the session&lt;/strong&gt;, my workflow with &lt;em&gt;AI&lt;/em&gt; has again changed so much that maybe it’s time for a &lt;strong&gt;part 2 of the blog post&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;The &lt;strong&gt;&lt;em&gt;“fancy autocomplete”&lt;/em&gt;&lt;/strong&gt; story started at the beginning of the year with the &lt;strong&gt;&lt;em&gt;&lt;a href=&quot;/blog/2025/copilot-fancy-autocomplete/&quot;&gt;part 1 blog post&lt;/a&gt;&lt;/em&gt;&lt;/strong&gt;. All the points from the blog post then matured into a &lt;strong&gt;&lt;em&gt;&lt;a href=&quot;https://www.youtube.com/watch?v=zTejuQzYAb8&quot;&gt;session at BC TechDays 2025&lt;/a&gt;&lt;/em&gt;&lt;/strong&gt;, which is also now available online.&lt;/p&gt;

&lt;div class=&quot;responsive-iframe&quot;&gt;
&lt;iframe src=&quot;https://www.youtube.com/embed/zTejuQzYAb8?si=NjXpEsyXi2BPUAqW&quot; title=&quot;YouTube video player&quot; frameborder=&quot;0&quot; allow=&quot;accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share&quot; referrerpolicy=&quot;strict-origin-when-cross-origin&quot; allowfullscreen=&quot;&quot;&gt;&lt;/iframe&gt;
&lt;/div&gt;
&lt;p&gt;&lt;br /&gt;&lt;/p&gt;

&lt;p&gt;The session covered some &lt;strong&gt;best practices&lt;/strong&gt; to get the best results out of the &lt;em&gt;fancy autocomplete&lt;/em&gt;, but then quickly turned towards &lt;strong&gt;other capabilities&lt;/strong&gt;. &lt;em&gt;Copilot Edits&lt;/em&gt; take the center stage, as the next step, and we later progress into &lt;strong&gt;&lt;em&gt;agent mode&lt;/em&gt;&lt;/strong&gt;, and compare &lt;em&gt;VS Code&lt;/em&gt; to &lt;em&gt;Cursor&lt;/em&gt;.&lt;/p&gt;

&lt;p&gt;If you haven’t seen the session yet, or if you’re at the start of your &lt;em&gt;Copilot journey&lt;/em&gt;, I’d suggest you first take the &lt;strong&gt;90 minutes&lt;/strong&gt; to watch the session. &lt;strong&gt;This blog post picks up where the session ended.&lt;/strong&gt;&lt;/p&gt;

&lt;hr /&gt;

&lt;p&gt;The most significant change is that, during the session, I take a stance that &lt;strong&gt;&lt;em&gt;Agent mode is nice, but Edits is where our focus should be.&lt;/em&gt;&lt;/strong&gt; Realistically, during the last 3 months, I’ve &lt;strong&gt;exclusively used only Agent mode&lt;/strong&gt; for both &lt;em&gt;vibe-coded projects&lt;/em&gt; and &lt;em&gt;AL tasks&lt;/em&gt;, and I don’t even look at &lt;em&gt;Edits&lt;/em&gt; anymore.&lt;/p&gt;

&lt;p&gt;I do believe &lt;em&gt;Edits&lt;/em&gt; will eventually be removed. &lt;em&gt;Cursor&lt;/em&gt; has already removed its &lt;strong&gt;&lt;em&gt;“manual”&lt;/em&gt;&lt;/strong&gt; mode, which was the &lt;em&gt;Edit equivalent&lt;/em&gt;. But for the time being, I think they can still have a purpose. If you’re &lt;strong&gt;entirely new to the Copilot journey&lt;/strong&gt;, especially if you’re a &lt;em&gt;skeptic&lt;/em&gt;, it can be a nice, &lt;strong&gt;slower introduction&lt;/strong&gt; to how &lt;em&gt;AI&lt;/em&gt; can help read and modify multiple files at scale, compared to the &lt;em&gt;Agent mode&lt;/em&gt;, which will have more &lt;strong&gt;&lt;em&gt;“a mind of its own”&lt;/em&gt;&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;I already introduced &lt;em&gt;Agents&lt;/em&gt; during the session, but here it comes again. &lt;strong&gt;&lt;em&gt;Agent mode is similar to Edit mode&lt;/em&gt;&lt;/strong&gt;, except that it tries to &lt;strong&gt;find its own context&lt;/strong&gt; for the changes it needs to make, and it has access to &lt;strong&gt;&lt;em&gt;“tools”&lt;/em&gt;&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol2/context.png&quot; alt=&quot;Agent Context&quot; /&gt;&lt;/p&gt;

&lt;p&gt;I made a &lt;strong&gt;big deal&lt;/strong&gt; in the session about &lt;strong&gt;&lt;em&gt;taking control of your context&lt;/em&gt;&lt;/strong&gt;. But it feels like that’s &lt;strong&gt;less important now&lt;/strong&gt;. If you have specific files that you want the agent to look at, by all means, pull them in. But also rest assured that it will do a &lt;strong&gt;pretty good job&lt;/strong&gt; at finding the files on its own:&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol2/context-2.png&quot; alt=&quot;Agent Context 2&quot; /&gt;&lt;/p&gt;

&lt;p&gt;It’s able to do that because it has &lt;strong&gt;access to tools&lt;/strong&gt;. It’s no longer just &lt;em&gt;reading content&lt;/em&gt; and &lt;em&gt;generating new tokens&lt;/em&gt;. It can &lt;strong&gt;take action within your IDE&lt;/strong&gt;, like &lt;em&gt;search for files&lt;/em&gt;, &lt;em&gt;run commands&lt;/em&gt;, &lt;em&gt;do web search&lt;/em&gt;, &lt;em&gt;check errors and warnings&lt;/em&gt;, …&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol2/tools.png&quot; alt=&quot;Agent Tools&quot; /&gt;&lt;/p&gt;

&lt;p&gt;We’ll return to &lt;strong&gt;tools&lt;/strong&gt; when we discuss &lt;em&gt;MCP Servers&lt;/em&gt;, but for now, know that this is the &lt;strong&gt;&lt;em&gt;“secret sauce”&lt;/em&gt;&lt;/strong&gt; behind the power of &lt;em&gt;Agent mode&lt;/em&gt;.&lt;/p&gt;

&lt;hr /&gt;

&lt;h2 id=&quot;agent-settings-for-best-results&quot;&gt;Agent Settings for Best Results&lt;/h2&gt;

&lt;p&gt;To have the &lt;strong&gt;best results with Agents&lt;/strong&gt;, here are two settings you should enable:&lt;/p&gt;

&lt;h3 id=&quot;todo-list-tool&quot;&gt;Todo List Tool&lt;/h3&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol2/todo-setting.png&quot; alt=&quot;Todo Setting&quot; /&gt;&lt;/p&gt;

&lt;p&gt;This tool will allow the agent to &lt;strong&gt;create its own to-do list&lt;/strong&gt; as it works through a more extensive task. It helps the agent &lt;strong&gt;&lt;em&gt;“remember”&lt;/em&gt;&lt;/strong&gt; what it has to implement throughout the length of a more extended conversation. Here’s how it looks:&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol2/todo-agent.png&quot; alt=&quot;Todo Agent&quot; /&gt;&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol2/todo-example.png&quot; alt=&quot;Todo Example&quot; /&gt;&lt;/p&gt;

&lt;p&gt;As it goes through the individual &lt;em&gt;Todo items&lt;/em&gt;, it &lt;strong&gt;marks them as completed&lt;/strong&gt;, helping it maintain focus on the current task.&lt;/p&gt;

&lt;h3 id=&quot;max-requests-setting&quot;&gt;Max Requests Setting&lt;/h3&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol2/max-requests.png&quot; alt=&quot;Max Requests&quot; /&gt;&lt;/p&gt;

&lt;p&gt;By default, the value is set to &lt;strong&gt;15&lt;/strong&gt; (or at least it was, when I changed it for myself). After &lt;em&gt;15 steps&lt;/em&gt;, the agent stops and asks the user, &lt;strong&gt;&lt;em&gt;“Do you want to continue or stop the process?”&lt;/em&gt;&lt;/strong&gt;. Fifteen steps is &lt;strong&gt;not a lot of steps&lt;/strong&gt;. I quickly grew annoyed at having to keep pressing &lt;em&gt;‘Continue’&lt;/em&gt; on larger tasks, so I changed mine to &lt;strong&gt;100&lt;/strong&gt;. I haven’t seen an interruption since, and I’m free to &lt;strong&gt;browse LinkedIn&lt;/strong&gt; as the agent implements larger items of my work. :P&lt;/p&gt;

&lt;hr /&gt;

&lt;h2 id=&quot;copilot-instructions&quot;&gt;Copilot Instructions&lt;/h2&gt;

&lt;p&gt;If you want to get the &lt;strong&gt;best results with the agent mode&lt;/strong&gt;, you should couple it with &lt;strong&gt;&lt;em&gt;copilot-instructions.md&lt;/em&gt;&lt;/strong&gt;. The instructions file will be &lt;strong&gt;appended to each request&lt;/strong&gt; you make to the agent. During the session, I explained how you can use this file as a &lt;strong&gt;&lt;em&gt;“memory” file&lt;/em&gt;&lt;/strong&gt;, to make sure the agent codes the way you would. Do you omit spaces in &lt;em&gt;object and field names&lt;/em&gt;, but agents use them? &lt;strong&gt;Put it in the file.&lt;/strong&gt; Do you want to have explicit parameters with &lt;em&gt;internal functions&lt;/em&gt; (Insert, Modify, Delete), but the agent skips them? &lt;strong&gt;Put it in the file.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;When creating these files, you don’t have to &lt;strong&gt;start from scratch&lt;/strong&gt;. &lt;strong&gt;&lt;em&gt;&lt;a href=&quot;https://www.linkedin.com/in/dmitrykatson&quot;&gt;Dmitry Katson&lt;/a&gt;&lt;/em&gt;&lt;/strong&gt; shared his &lt;strong&gt;&lt;em&gt;&lt;a href=&quot;https://alguidelines.dev/docs/vibe-coding/&quot;&gt;instruction files&lt;/a&gt;&lt;/em&gt;&lt;/strong&gt; with the community. I keep mine &lt;strong&gt;&lt;em&gt;&lt;a href=&quot;https://github.com/tinestaric/BCExamples/blob/Master/.github/copilot-instructions.md&quot;&gt;here&lt;/a&gt;&lt;/em&gt;&lt;/strong&gt;, but they’re far less extensive, as I’ve generally had &lt;strong&gt;good experience&lt;/strong&gt; with the proposed code, with only minor adjustments needed that I keep in the instructions file.&lt;/p&gt;

&lt;p&gt;However, something that I haven’t mentioned in the session, nor captured in the above instructions file, that I think should be part of &lt;strong&gt;every single repository&lt;/strong&gt; is a &lt;strong&gt;&lt;em&gt;project overview&lt;/em&gt;&lt;/strong&gt;. Every project should have an explanation of what it’s about, what the architecture looks like, and the main patterns used. And even there, you don’t have to &lt;strong&gt;start from scratch&lt;/strong&gt;. A few releases ago, the &lt;em&gt;VSC team&lt;/em&gt; added the action to &lt;strong&gt;generate the project overview instructions&lt;/strong&gt;:&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol2/generate-instructions.png&quot; alt=&quot;Generate Instructions&quot; /&gt;&lt;/p&gt;

&lt;p&gt;When executed, it will prompt the agent with a &lt;strong&gt;preset prompt&lt;/strong&gt; to scan the workspace and create (or update) the instructions file with the project overview:&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol2/instructions-prompt.png&quot; alt=&quot;Instructions Prompt&quot; /&gt;&lt;/p&gt;

&lt;p&gt;And here are some of the results&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol2/instructions-example.png&quot; alt=&quot;Instructions Example&quot; /&gt;&lt;/p&gt;

&lt;p&gt;Considering how &lt;strong&gt;easy it is&lt;/strong&gt; to include an overview, I don’t see a reason why you wouldn’t go and include it in your repos &lt;strong&gt;&lt;em&gt;right now&lt;/em&gt;&lt;/strong&gt;. :P&lt;/p&gt;

&lt;hr /&gt;

&lt;h2 id=&quot;models&quot;&gt;Models&lt;/h2&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol2/models.jpg&quot; alt=&quot;Models&quot; /&gt;&lt;/p&gt;

&lt;p&gt;Next, I want to touch on the topic of &lt;strong&gt;models&lt;/strong&gt; again. My stance remains the same; &lt;strong&gt;&lt;em&gt;Claude models appear to be the best for AL&lt;/em&gt;&lt;/strong&gt;. However, we’ve recently received &lt;strong&gt;two others&lt;/strong&gt; that I sometimes like to switch to, especially when &lt;em&gt;Claude&lt;/em&gt; appears to &lt;strong&gt;&lt;em&gt;get drunk&lt;/em&gt;&lt;/strong&gt; and proposes &lt;em&gt;stupid solutions&lt;/em&gt; (this one was suggested to &lt;em&gt;code hand-crafter&lt;/em&gt; &lt;strong&gt;&lt;em&gt;&lt;a href=&quot;https://www.linkedin.com/in/ajkauffmann&quot;&gt;AJ Kauffmann&lt;/a&gt;&lt;/em&gt;&lt;/strong&gt;):&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol2/dumb-agent.jpg&quot; alt=&quot;Dumb Agent&quot; /&gt;&lt;/p&gt;

&lt;p&gt;The &lt;strong&gt;models I like for AL&lt;/strong&gt; are:&lt;/p&gt;
&lt;ul&gt;
  &lt;li&gt;&lt;strong&gt;&lt;em&gt;Claude Sonnet 4&lt;/em&gt;&lt;/strong&gt; – My default, &lt;strong&gt;best results&lt;/strong&gt;, most consistent, overall the &lt;strong&gt;best experience&lt;/strong&gt;.&lt;/li&gt;
  &lt;li&gt;&lt;strong&gt;&lt;em&gt;GPT-5&lt;/em&gt;&lt;/strong&gt; – After poor &lt;em&gt;AL results&lt;/em&gt; with &lt;em&gt;4o&lt;/em&gt;, &lt;em&gt;4.1&lt;/em&gt;, &lt;em&gt;o3&lt;/em&gt;, &lt;em&gt;4.5&lt;/em&gt;, this is the &lt;strong&gt;first GPT model&lt;/strong&gt; that I can easily trust to get things right. I like to use it as an alternative to &lt;em&gt;Sonnet&lt;/em&gt; when I have tasks that require &lt;strong&gt;more of an analysis&lt;/strong&gt; of a codebase. Apparently, &lt;em&gt;Copilot&lt;/em&gt; was also tweaked to work better with this model, as evidenced by these &lt;strong&gt;two settings&lt;/strong&gt; I found (but I haven’t noticed too much of a difference)&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol2/gpt-5-settings.png&quot; alt=&quot;GPT-5 Settings&quot; /&gt;&lt;/p&gt;

&lt;ul&gt;
  &lt;li&gt;&lt;strong&gt;&lt;em&gt;Grok Code Fast 1&lt;/em&gt;&lt;/strong&gt; – This is the only &lt;strong&gt;&lt;em&gt;“included”&lt;/em&gt;&lt;/strong&gt; model (meaning &lt;em&gt;“free”&lt;/em&gt;) that I like to use with &lt;em&gt;AL&lt;/em&gt;. It’s &lt;strong&gt;faster&lt;/strong&gt; than the other two, and it does the work without all the meaningless tokens for &lt;strong&gt;&lt;em&gt;“You’re absolutely right!”&lt;/em&gt;&lt;/strong&gt;. It, however, &lt;strong&gt;doesn’t handle picture input&lt;/strong&gt;. Less relevant for &lt;em&gt;AL&lt;/em&gt;, but worth keeping in mind.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Still remember that if you don’t have any of these models available in your model selector, you (or your organization admin) have to enable them in &lt;strong&gt;&lt;em&gt;&lt;a href=&quot;https://github.com/settings/copilot/features&quot;&gt;GitHub Settings&lt;/a&gt;&lt;/em&gt;&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;And if you’re of the &lt;strong&gt;exploratory type&lt;/strong&gt; when it comes to models, you absolutely have to give &lt;strong&gt;&lt;em&gt;&lt;a href=&quot;https://www.linkedin.com/in/stefano-demiliani&quot;&gt;Stefano Demiliani&lt;/a&gt;&lt;/em&gt;&lt;/strong&gt; a follow. He &lt;strong&gt;tests each new model with AL&lt;/strong&gt; and reports his impression on all social media. I only touch models that &lt;strong&gt;he approves&lt;/strong&gt;. :P&lt;/p&gt;

&lt;hr /&gt;

&lt;h2 id=&quot;mcp-servers&quot;&gt;MCP Servers&lt;/h2&gt;

&lt;p&gt;Lastly, let’s talk about the &lt;strong&gt;&lt;em&gt;MCP servers&lt;/em&gt;&lt;/strong&gt;. They’ve been &lt;strong&gt;all around the news&lt;/strong&gt; for a few months now, but I think we’re right at the brink of getting a &lt;strong&gt;&lt;em&gt;“perfect bouquet of MCPs”&lt;/em&gt;&lt;/strong&gt; to use for &lt;em&gt;AL development&lt;/em&gt;.&lt;/p&gt;

&lt;p&gt;Before I jump into my current recommendations, I want to provide a &lt;strong&gt;short explanation&lt;/strong&gt; of what an &lt;em&gt;MCP server&lt;/em&gt; is. I’ll use the same analogy I used in the recent session for our partner.&lt;/p&gt;

&lt;p&gt;Everyone understands the concept of &lt;strong&gt;&lt;em&gt;VSC extensions&lt;/em&gt;&lt;/strong&gt;. We have this &lt;em&gt;vanilla development environment&lt;/em&gt; that allows us to write code. Still, if we want to add any &lt;strong&gt;additional functionalities&lt;/strong&gt; to it (like support for &lt;em&gt;AL&lt;/em&gt;), we need to install extensions for it.&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol2/vsc-extension.png&quot; alt=&quot;VSC Extension&quot; /&gt;&lt;/p&gt;

&lt;p&gt;But an extension that’s built for &lt;em&gt;VSC&lt;/em&gt; will &lt;strong&gt;not work&lt;/strong&gt; for &lt;em&gt;Visual Studio&lt;/em&gt; or &lt;em&gt;JetBrains&lt;/em&gt;.&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol2/vsc-extension-error.png&quot; alt=&quot;VSC Extension Error&quot; /&gt;&lt;/p&gt;

&lt;p&gt;The same story applies to &lt;em&gt;Copilot&lt;/em&gt; and other &lt;strong&gt;&lt;em&gt;AI-assisted development tools&lt;/em&gt;&lt;/strong&gt;. Remember that the &lt;strong&gt;true power&lt;/strong&gt; of the &lt;em&gt;GitHub Copilot Agent mode&lt;/em&gt; comes from its tools:&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol2/tools.png&quot; alt=&quot;Tools&quot; /&gt;&lt;/p&gt;

&lt;p&gt;But if we create an additional tool for &lt;em&gt;GitHub Copilot&lt;/em&gt;, it won’t work with tools like &lt;strong&gt;&lt;em&gt;Claude Code&lt;/em&gt;&lt;/strong&gt;, &lt;em&gt;Codex&lt;/em&gt;, &lt;em&gt;Aider&lt;/em&gt;, and others.&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol2/copilot-extension.png&quot; alt=&quot;Copilot Extension&quot; /&gt;&lt;/p&gt;

&lt;p&gt;And that’s what &lt;strong&gt;&lt;em&gt;MCP (Model Context Protocol)&lt;/em&gt;&lt;/strong&gt; is here to solve. If we build our &lt;em&gt;Copilot extension&lt;/em&gt; in a way that it’s &lt;strong&gt;&lt;em&gt;MCP-compliant&lt;/em&gt;&lt;/strong&gt;, then &lt;strong&gt;any tool that supports MCP&lt;/strong&gt; can make use of our additional tools.&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol2/mcp-extension.png&quot; alt=&quot;MCP Extension&quot; /&gt;&lt;/p&gt;

&lt;p&gt;Regardless of whether the tools are added through an &lt;em&gt;extension&lt;/em&gt; or through an &lt;em&gt;MCP server&lt;/em&gt;, their &lt;strong&gt;behaviour remains the same&lt;/strong&gt;. It’s just a protocol of how the tools &lt;strong&gt;&lt;em&gt;“plug in”&lt;/em&gt;&lt;/strong&gt; to the client.&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol2/mcp-examples.png&quot; alt=&quot;MCP Examples&quot; /&gt;&lt;/p&gt;

&lt;p&gt;A lot of people have used the &lt;strong&gt;&lt;em&gt;“USB-C”&lt;/em&gt;&lt;/strong&gt; analogy in the past, and it does fit. Let’s say you have a printer that plugs into your PC via &lt;em&gt;USB-C&lt;/em&gt;. The &lt;strong&gt;&lt;em&gt;“logic”&lt;/em&gt;&lt;/strong&gt; is in the printer, not in the &lt;em&gt;USB-C connector&lt;/em&gt;. The connector is just a connector. But it allows us to &lt;strong&gt;plug the printer into anything&lt;/strong&gt; that has a &lt;em&gt;USB-C hole&lt;/em&gt;. The same applies to &lt;em&gt;MCPs&lt;/em&gt;. The &lt;strong&gt;logic is in the “extension”&lt;/strong&gt;; &lt;em&gt;MCP&lt;/em&gt; ensures that we can plug our logic into &lt;strong&gt;anything that has an MCP hole&lt;/strong&gt;. :D&lt;/p&gt;

&lt;hr /&gt;

&lt;p&gt;Okay, so what &lt;strong&gt;&lt;em&gt;MCPs should you try out&lt;/em&gt;&lt;/strong&gt; for &lt;em&gt;AL development&lt;/em&gt;?&lt;/p&gt;

&lt;ul&gt;
  &lt;li&gt;&lt;strong&gt;&lt;em&gt;GitHub / Azure DevOps&lt;/em&gt;&lt;/strong&gt; – depending on what you use for issue tracking, these are an &lt;strong&gt;obvious choice&lt;/strong&gt;. It enables the agent to &lt;em&gt;read and update tasks&lt;/em&gt;, as well as handle &lt;em&gt;pipelines&lt;/em&gt;. &lt;strong&gt;Very useful&lt;/strong&gt;.&lt;/li&gt;
  &lt;li&gt;&lt;strong&gt;&lt;em&gt;Microsoft Docs&lt;/em&gt;&lt;/strong&gt; – I’ve had several tasks where the agent surprised me with &lt;strong&gt;&lt;em&gt;“let me search the documentation”&lt;/em&gt;&lt;/strong&gt;, found the &lt;em&gt;BC related articles&lt;/em&gt;, and continued with &lt;strong&gt;much better context&lt;/strong&gt;.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;These are the &lt;strong&gt;easy-to-install MCPs&lt;/strong&gt; that you can find by opening the &lt;strong&gt;&lt;em&gt;MCP library&lt;/em&gt;&lt;/strong&gt; from &lt;em&gt;VSC&lt;/em&gt;.&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol2/browser-mcp.png&quot; alt=&quot;Browser MCP&quot; /&gt;&lt;/p&gt;

&lt;hr /&gt;

&lt;p&gt;A worthwhile &lt;strong&gt;&lt;em&gt;non-MCP recommendation&lt;/em&gt;&lt;/strong&gt; is something you might already have in your toolbox: &lt;strong&gt;&lt;em&gt;NAB AL Tools&lt;/em&gt;&lt;/strong&gt;. &lt;strong&gt;&lt;em&gt;&lt;a href=&quot;https://www.linkedin.com/in/johanneswikman/&quot;&gt;Johannes&lt;/a&gt;&lt;/em&gt;&lt;/strong&gt; added &lt;em&gt;Copilot tools&lt;/em&gt; to his already existing extension, and if you’re working with &lt;strong&gt;translations&lt;/strong&gt;, the following blog is a &lt;strong&gt;&lt;em&gt;&lt;a href=&quot;https://www.linkedin.com/pulse/language-model-tools-ai-translations-johannes-wikman-qiebf/?trackingId=BI%2BPcryjdFcmxbaQ3u5O5Q%3D%3D&quot;&gt;MUST READ&lt;/a&gt;&lt;/em&gt;&lt;/strong&gt;.&lt;/p&gt;

&lt;hr /&gt;

&lt;p&gt;But back to &lt;em&gt;MCPs&lt;/em&gt;, the following ones will need a &lt;strong&gt;bit more manual work&lt;/strong&gt;, but everything is explained in the readme files:&lt;/p&gt;
&lt;ul&gt;
  &lt;li&gt;&lt;strong&gt;&lt;em&gt;&lt;a href=&quot;https://github.com/StefanMaron/AL-Dependency-MCP-Server&quot;&gt;Al-symbols-mcp&lt;/a&gt;&lt;/em&gt;&lt;/strong&gt; - Created by &lt;strong&gt;&lt;em&gt;&lt;a href=&quot;https://www.linkedin.com/in/stefan-maron-709928206/&quot;&gt;Stefan Maron&lt;/a&gt;&lt;/em&gt;&lt;/strong&gt;, it allows the agent to &lt;strong&gt;look into the dependencies&lt;/strong&gt; for additional context. Usually, when working on an app, the agent can only see the code from the workspace. This &lt;em&gt;MCP&lt;/em&gt; adds tools that can give it access to &lt;strong&gt;objects that are in the .alpackages folder&lt;/strong&gt;. It’s &lt;strong&gt;brilliant for PTE apps&lt;/strong&gt; where most of the context is in the dependencies—more information in the &lt;em&gt;ReadMe file&lt;/em&gt; in the repo.&lt;/li&gt;
  &lt;li&gt;&lt;strong&gt;&lt;em&gt;&lt;a href=&quot;https://www.npmjs.com/package/bc-code-intelligence-mcp&quot;&gt;Bc-code-intelligence-mcp&lt;/a&gt;&lt;/em&gt;&lt;/strong&gt; - This one is from &lt;strong&gt;&lt;em&gt;&lt;a href=&quot;https://www.linkedin.com/in/jeremyvyska/&quot;&gt;Jeremy Vyska&lt;/a&gt;&lt;/em&gt;&lt;/strong&gt;, another &lt;strong&gt;must-follow person&lt;/strong&gt; if you’re into the &lt;em&gt;Copilot story&lt;/em&gt;. It’s hard to believe &lt;em&gt;Jeremy&lt;/em&gt; was in the &lt;strong&gt;&lt;em&gt;“autocomplete”&lt;/em&gt;&lt;/strong&gt; camp 6 months ago, because he has already done &lt;strong&gt;so much&lt;/strong&gt; to push the &lt;em&gt;Copilot story&lt;/em&gt; further for &lt;em&gt;AL devs&lt;/em&gt;! Anyway, this &lt;em&gt;MCP&lt;/em&gt; brings &lt;strong&gt;&lt;em&gt;A TON of AL dev guidelines&lt;/em&gt;&lt;/strong&gt; to your agent so it writes code as it should be written. It’s hard to explain everything this &lt;em&gt;MCP&lt;/em&gt; does, so it’s best if you check &lt;strong&gt;&lt;em&gt;&lt;a href=&quot;https://nubimancy.com/2025/09/22/bc-code-intelligence-launch-announcement/&quot;&gt;Jeremy’s recent blog post&lt;/a&gt;&lt;/em&gt;&lt;/strong&gt;.&lt;/li&gt;
&lt;/ul&gt;

&lt;hr /&gt;

&lt;p&gt;The last &lt;em&gt;MCP&lt;/em&gt; recommendations are a bit more &lt;strong&gt;experimental&lt;/strong&gt;. I have not spent enough time with them, but I have a &lt;strong&gt;strong feeling&lt;/strong&gt; they’re going to be very instrumental in the future:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;&lt;em&gt;&lt;a href=&quot;https://github.com/oraios/serena&quot;&gt;Serena&lt;/a&gt;&lt;/em&gt;&lt;/strong&gt; - By default, the agent navigates through your codebase using &lt;em&gt;text search&lt;/em&gt;. &lt;em&gt;Serena&lt;/em&gt; enables it to do &lt;strong&gt;semantic retrieval&lt;/strong&gt;, navigate through the codebase on the &lt;em&gt;symbol level&lt;/em&gt;, and exploit &lt;strong&gt;relational structure&lt;/strong&gt;. Usually, these advanced tools would be reserved for &lt;strong&gt;&lt;em&gt;“mature”&lt;/em&gt;&lt;/strong&gt; languages like &lt;em&gt;Python&lt;/em&gt;, &lt;em&gt;C#&lt;/em&gt;, or &lt;em&gt;JavaScript&lt;/em&gt;, but thanks to the &lt;strong&gt;&lt;em&gt;“madman”&lt;/em&gt;&lt;/strong&gt; named &lt;strong&gt;&lt;em&gt;&lt;a href=&quot;https://www.linkedin.com/in/torben-l%C3%B8kke-leth/&quot;&gt;Torben Leth&lt;/a&gt;&lt;/em&gt;&lt;/strong&gt;, they have just received a &lt;strong&gt;&lt;em&gt;&lt;a href=&quot;https://github.com/oraios/serena/pull/593&quot;&gt;PR&lt;/a&gt;&lt;/em&gt;&lt;/strong&gt; with &lt;em&gt;AL support&lt;/em&gt;. I use &lt;em&gt;madman&lt;/em&gt; in a &lt;strong&gt;very complimentary tone&lt;/strong&gt;. :D &lt;em&gt;Torben&lt;/em&gt; is one of those people who gets involved in projects that I’d consider &lt;strong&gt;impossible&lt;/strong&gt; (like &lt;strong&gt;&lt;em&gt;&lt;a href=&quot;https://www.linkedin.com/posts/torben-l%C3%B8kke-leth_emea2025-msdyn365bc-activity-7367584117506465795-4_Uy?utm_source=social_share_send&amp;amp;utm_medium=member_desktop_web&amp;amp;rcm=ACoAACLigK8B1d5zenhN3aNEAnW6bNGFZZNzj_E&quot;&gt;bringing BC to Linux&lt;/a&gt;&lt;/em&gt;&lt;/strong&gt;).&lt;/p&gt;

&lt;p&gt;He also just published the &lt;em&gt;MCP&lt;/em&gt; for the &lt;strong&gt;&lt;em&gt;AL Object ID Ninja&lt;/em&gt;&lt;/strong&gt; made by &lt;strong&gt;&lt;em&gt;&lt;a href=&quot;https://www.linkedin.com/in/vjeko/&quot;&gt;Vjeko&lt;/a&gt;&lt;/em&gt;&lt;/strong&gt;. Finally, agents will &lt;strong&gt;know which ID to propose&lt;/strong&gt;, but I have yet to test this one out: &lt;strong&gt;&lt;em&gt;&lt;a href=&quot;https://github.com/SShadowS/al-objid-mcp-server&quot;&gt;al-objid-mcp-server&lt;/a&gt;&lt;/em&gt;&lt;/strong&gt;.&lt;/p&gt;

&lt;hr /&gt;

&lt;blockquote&gt;
  &lt;p&gt;&lt;strong&gt;Warning:&lt;/strong&gt; Don’t go crazy installing &lt;strong&gt;&lt;em&gt;all the MCPs&lt;/em&gt;&lt;/strong&gt; you can find. Every added tool will &lt;strong&gt;eat away at the context window&lt;/strong&gt; of an agent, and there will be &lt;strong&gt;&lt;em&gt;less space for context&lt;/em&gt;&lt;/strong&gt; about your project or the goal. &lt;strong&gt;&lt;em&gt;&lt;a href=&quot;https://www.linkedin.com/in/stefano-demiliani&quot;&gt;Stefano&lt;/a&gt;&lt;/em&gt;&lt;/strong&gt; described it perfectly in his recent &lt;strong&gt;&lt;em&gt;&lt;a href=&quot;https://demiliani.com/2025/09/04/model-context-protocol-and-the-too-many-tools-problem/&quot;&gt;blog post&lt;/a&gt;&lt;/em&gt;&lt;/strong&gt;.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;&lt;em&gt;VS Code&lt;/em&gt; will warn you if you’re trying to fit &lt;strong&gt;more than 128 tools&lt;/strong&gt; in your prompt, but even &lt;em&gt;128&lt;/em&gt; is already a &lt;strong&gt;huge number&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/fancyautocompletevol2/too-many-tools.png&quot; alt=&quot;Too Many Tools&quot; /&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Azure DevOps MCP&lt;/em&gt; alone comes with &lt;strong&gt;70 new tools&lt;/strong&gt;, and do you really need to have all the actions related to &lt;em&gt;pipelines&lt;/em&gt; enabled when your usual workflow consists of &lt;strong&gt;&lt;em&gt;reading and maybe updating work items?&lt;/em&gt;&lt;/strong&gt; I think not.&lt;/p&gt;

&lt;hr /&gt;

&lt;p&gt;Was this too long? Nah, I had to put all the current ideas into &lt;em&gt;vol.2&lt;/em&gt;. I already know there’s going to be a &lt;strong&gt;&lt;em&gt;vol. 3&lt;/em&gt;&lt;/strong&gt; soon. The next area worth mentioning is going to be the &lt;strong&gt;background agents&lt;/strong&gt;, the ones we no longer oversee. I also think we’ll see &lt;strong&gt;many more MCPs&lt;/strong&gt; in the next few months, until they all consolidate into an &lt;strong&gt;&lt;em&gt;&lt;a href=&quot;https://marketplace.visualstudio.com/items?itemName=waldo.al-extension-pack&quot;&gt;AL Extension Pack&lt;/a&gt;&lt;/em&gt;&lt;/strong&gt; equivalent of &lt;em&gt;MCP servers&lt;/em&gt;. I think it’s also time to &lt;strong&gt;revisit the topic of code review&lt;/strong&gt;, this time through an &lt;em&gt;AI-assisted lens&lt;/em&gt;.&lt;/p&gt;

&lt;p&gt;Let me know your thoughts on any of the social media, and see you on the next post!&lt;/p&gt;

</description>
        <pubDate>Sun, 21 Sep 2025 10:00:00 +0200</pubDate>
        <link>https://tine.staric.net/blog/2025/copilot-fancy-autocomplete-vol2/</link>
        <guid isPermaLink="true">https://tine.staric.net/blog/2025/copilot-fancy-autocomplete-vol2/</guid>
        
        
        <category>al</category>
        
      </item>
    
      <item>
        <title>100 Areopa Webinars Visualized</title>
        <description>&lt;p&gt;&lt;strong&gt;&lt;a href=&quot;https://areopa.academy/&quot;&gt;Areopa Webinars&lt;/a&gt;&lt;/strong&gt; just hit number &lt;strong&gt;&lt;em&gt;100!&lt;/em&gt;&lt;/strong&gt; And I used this milestone as another &lt;strong&gt;visualization side-project.&lt;/strong&gt; :P&lt;/p&gt;

&lt;p&gt;If you don’t know &lt;strong&gt;Areopa&lt;/strong&gt;, the short description of it is a &lt;em&gt;YouTube channel&lt;/em&gt; of BC-related webinars that have been going strong most Mondays since &lt;strong&gt;&lt;em&gt;2019!&lt;/em&gt;&lt;/strong&gt; &lt;a href=&quot;https://www.linkedin.com/in/lvanvugt/&quot;&gt;Luc van Vugt&lt;/a&gt; leads it, but the &lt;strong&gt;&lt;em&gt;whole community&lt;/em&gt;&lt;/strong&gt; contributed to the content! Luc, if you’re reading this, you really do deserve a shoutout… You didn’t only create this huge (and still growing) collection of BC knowledge, but you’ve also setup a platform where &lt;strong&gt;&lt;em&gt;anyone&lt;/em&gt;&lt;/strong&gt; can come and share what they know with &lt;strong&gt;&lt;em&gt;everyone&lt;/em&gt;&lt;/strong&gt;!&lt;/p&gt;

&lt;p&gt;Here’s why I care so much about Areopa. I’ve been very lucky to be in a position, where I get to travel to multiple conferences every year, and I listen to so many amazing sessions. It’s easy to &lt;strong&gt;stay &lt;em&gt;“up to date”&lt;/em&gt;&lt;/strong&gt; when I hear about best practices all the time, directly from the industry’s best. But &lt;strong&gt;&lt;em&gt;not everyone gets that chance&lt;/em&gt;&lt;/strong&gt;. And here’s where Areopa really excels, it brings that same level of high-quality sessions to &lt;strong&gt;&lt;em&gt;everyone! For free! Forever!&lt;/em&gt;&lt;/strong&gt; Conference sessions are great, but we &lt;strong&gt;forget&lt;/strong&gt; most of what we hear the very &lt;strong&gt;next day.&lt;/strong&gt; Areopa makes it available forever. You can always get back to a certain topic once it’s relevant to you again.&lt;/p&gt;

&lt;hr /&gt;

&lt;p&gt;I didn’t want to just give praise to Areopa though. As I mentioned at the beginning, I used this opportunity as another &lt;strong&gt;visualization side project.&lt;/strong&gt; Back in November, after &lt;strong&gt;Ignite&lt;/strong&gt;, I came across &lt;a href=&quot;https://markmap.js.org/full#?d=gist:38520ad7809001358250f62e88695a80:2024-Ignite-Book-of-News.md&quot;&gt;this visualization&lt;/a&gt; by &lt;a href=&quot;https://www.linkedin.com/in/nityan/&quot;&gt;Nitya Narasimhan&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;&lt;img src=&quot;/images/areopavisualized/ignite.png&quot; alt=&quot;Ignite sessions visualized&quot; /&gt;&lt;/p&gt;

&lt;p&gt;She grouped all Ignite sessions into groups and visualized them. I was so amazed with the concept that I knew I needed to find some time to play around with &lt;strong&gt;Markmap&lt;/strong&gt;. My initial idea was (and still is) to make a similar visualization for all sessions at &lt;strong&gt;Directions EMEA&lt;/strong&gt;, so I can have an easier time deciding what I’d like to listen to. I mean, it’s super hard to keep track of &lt;strong&gt;~200 sessions&lt;/strong&gt; otherwise…&lt;/p&gt;

&lt;p&gt;I’ve started being &lt;strong&gt;more involved with Areopa&lt;/strong&gt; within the past few months. &lt;em&gt;Finding speakers, moderating webinars, and presenting my own topics,&lt;/em&gt; and as we were approaching this &lt;strong&gt;triple-digit number&lt;/strong&gt;, I realized I have no clue about half the videos available on the channel. I was thinking about how could we make this huge collection of videos &lt;strong&gt;more discoverable.&lt;/strong&gt; You can already see how this turned into a perfect opportunity to play with &lt;strong&gt;Markmap&lt;/strong&gt;. Below you’ll find the result. I invite you to explore the graph, and if any video sparks your interest, go ahead and watch it! :P&lt;/p&gt;

&lt;hr /&gt;

&lt;p&gt;But just before I let you go, I wanted to share a few &lt;strong&gt;“fun facts”&lt;/strong&gt; that I found as I was handling the data of the videos:&lt;/p&gt;

&lt;ul&gt;
  &lt;li&gt;First video’s publish date: Jul 7, 2019&lt;/li&gt;
  &lt;li&gt;Total number of unique speakers: 64&lt;/li&gt;
&lt;/ul&gt;

&lt;h3 id=&quot;top-5-most-frequent-presenters&quot;&gt;Top 5 most frequent presenters:&lt;/h3&gt;
&lt;ul&gt;
  &lt;li&gt;&lt;a href=&quot;https://www.linkedin.com/in/lvanvugt/&quot;&gt;Luc van Vugt&lt;/a&gt; – 13 webinars&lt;/li&gt;
  &lt;li&gt;&lt;a href=&quot;https://www.linkedin.com/in/tobiasfenster&quot;&gt;Tobias Fenster&lt;/a&gt; – 8 webinars&lt;/li&gt;
  &lt;li&gt;&lt;a href=&quot;https://www.linkedin.com/in/ajkauffmann&quot;&gt;Arend-Jan Kauffmann&lt;/a&gt; – 6 webinars&lt;/li&gt;
  &lt;li&gt;&lt;a href=&quot;https://www.linkedin.com/in/ericwauters&quot;&gt;waldo&lt;/a&gt; – 6 webinars&lt;/li&gt;
  &lt;li&gt;&lt;a href=&quot;https://www.linkedin.com/in/dmitrykatson&quot;&gt;Dmitry Katson&lt;/a&gt; – 5 webinars&lt;/li&gt;
&lt;/ul&gt;

&lt;h3 id=&quot;top-5-most-watched-webinars&quot;&gt;Top 5 most watched webinars:&lt;/h3&gt;
&lt;ul&gt;
  &lt;li&gt;&lt;a href=&quot;https://youtube.com/watch?v=mNgx1d7TAXQ&quot;&gt;Replicating Business Central data to Microsoft Dataverse (former CDS)&lt;/a&gt; (by &lt;a href=&quot;https://www.linkedin.com/in/andrey-baludin-9a7014191&quot;&gt;Andrey Baludin&lt;/a&gt;)&lt;/li&gt;
  &lt;li&gt;&lt;a href=&quot;https://youtube.com/watch?v=4X-tSjL0XaA&quot;&gt;Working with files in Dynamics 365 Business Central SaaS&lt;/a&gt; (by &lt;a href=&quot;https://www.linkedin.com/in/stefano-demiliani&quot;&gt;Stefano Demiliani&lt;/a&gt;)&lt;/li&gt;
  &lt;li&gt;&lt;a href=&quot;https://youtube.com/watch?v=Fjz9LgviV2Q&quot;&gt;Sync your Business Central data with Azure Data Lake&lt;/a&gt; (by &lt;a href=&quot;https://www.linkedin.com/in/bertverbeek&quot;&gt;Bert Verbeek&lt;/a&gt;)&lt;/li&gt;
  &lt;li&gt;&lt;a href=&quot;https://youtube.com/watch?v=3w018zjbUwQ&quot;&gt;Working with XML in AL&lt;/a&gt; (by &lt;a href=&quot;https://www.linkedin.com/in/ajkauffmann&quot;&gt;Arend-Jan Kauffmann&lt;/a&gt;)&lt;/li&gt;
  &lt;li&gt;&lt;a href=&quot;https://youtube.com/watch?v=yeKRuw9MtSQ&quot;&gt;OAuth authentication with Business Central APIs&lt;/a&gt; (by &lt;a href=&quot;https://www.linkedin.com/in/ajkauffmann&quot;&gt;Arend-Jan Kauffmann&lt;/a&gt;)&lt;/li&gt;
&lt;/ul&gt;

&lt;h3 id=&quot;most-common-topics&quot;&gt;Most common topics:&lt;/h3&gt;
&lt;ul&gt;
  &lt;li&gt;Test Automation: 16 videos&lt;/li&gt;
  &lt;li&gt;AL: 12 videos&lt;/li&gt;
  &lt;li&gt;DevOps: 12 videos&lt;/li&gt;
  &lt;li&gt;Functional: 12 videos&lt;/li&gt;
  &lt;li&gt;Power Platform: 7 videos&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Interested in what’s under these topics? Well then I won’t keep you waiting any longer, here’s the visualization. You can zoom in and out, move the graph around, and click on the nodes to expand or collapse them. &lt;strong&gt;&lt;em&gt;Click around, explore!&lt;/em&gt;&lt;/strong&gt;&lt;/p&gt;

&lt;!-- Load required libraries --&gt;
&lt;script src=&quot;https://cdn.jsdelivr.net/npm/d3@7&quot;&gt;&lt;/script&gt;

&lt;script src=&quot;https://cdn.jsdelivr.net/npm/markmap-view&quot;&gt;&lt;/script&gt;

&lt;script src=&quot;https://cdn.jsdelivr.net/npm/markmap-lib&quot;&gt;&lt;/script&gt;

&lt;!-- Mindmap container --&gt;
&lt;div id=&quot;mindmap&quot;&gt;
  &lt;svg style=&quot;width: 100%; height: 800px&quot;&gt;&lt;/svg&gt;
&lt;/div&gt;

&lt;!-- Initialize and render mindmap --&gt;
&lt;script&gt;
document.addEventListener(&apos;DOMContentLoaded&apos;, async () =&gt; {
  const { Markmap } = window.markmap;
  
  // Fetch markdown content straight from the repo (raw.githubusercontent.com sends CORS headers, no proxy needed)
  try {
    const response = await fetch(&apos;https://raw.githubusercontent.com/tinestaric/AreopaAnalyzer/main/data/videos_markmap.md&apos;);
    if (!response.ok) {
      throw new Error(&apos;Failed to fetch videos_markmap.md, status &apos; + response.status);
    }
    const markdown = await response.text();

    // Transform markdown to mindmap data
    const { Transformer } = window.markmap;
    const transformer = new Transformer();
    const { root } = transformer.transform(markdown);

    // Configure visualization options
    const { deriveOptions } = window.markmap;
    const jsonOptions = {
      initialExpandLevel: 2,
      colorFreezeLevel: 3,
      duration: 1000,
      spacingVertical: 10
    };
    const options = deriveOptions(jsonOptions);

    // Create and render visualization
    const svg = d3.select(&quot;#mindmap svg&quot;);
    const mm = Markmap.create(svg.node(), options, root);
  } catch (error) {
    console.error(&apos;Error fetching markmap data from release:&apos;, error);
    document.getElementById(&apos;mindmap&apos;).innerHTML = &apos;&lt;p style=&quot;text-align: center; color: #666; padding: 40px;&quot;&gt;Unable to load visualization. Please try again later.&lt;/p&gt;&apos;;
  }
});
&lt;/script&gt;

</description>
        <pubDate>Sun, 02 Feb 2025 09:00:00 +0100</pubDate>
        <link>https://tine.staric.net/blog/2025/areopa-visualized/</link>
        <guid isPermaLink="true">https://tine.staric.net/blog/2025/areopa-visualized/</guid>
        
        
        <category>al</category>
        
      </item>
    
  </channel>
</rss>
