The AL Coding Harness: Your Custom Dashboard

Note: This blog is a personal notebook. I write it to record what I work on, for my own reference, not to publish it on social media or to promote myself. If a post is still useful to you, I am glad, but that is not why it exists.

This post was published on June 29, 2026. Software versions, product features, links, and tool availability may have changed since then. Always check official documentation for the latest information. Some screenshots show small details. Zoom in on the page to see them clearly.

A language model provides text. It does not open your files, run your tests, or apply an edit. Another layer does that work and sends each result back to the model. The industry calls that layer the harness.

Anthropic used the word for the layer around Claude Code in September 2025. The VS Code team described theirs in May 2026, in The Coding Harness Behind GitHub Copilot in VS Code.

Update (2026-08-21): For a quick look at how a harness works, see Understanding Agent Harnesses on YouTube.

The engine, the chassis, and the dashboard

A car makes this structure easy to see. The model is the engine. The harness is the chassis around it. Your project files are the custom dashboard and the special equipment that you add for one track.

  1. The model, the engine. It reasons and writes the answer.
  2. The harness, the chassis. It remembers the context, gets the information, validates the results, and executes the changes.
  3. Your configuration, the dashboard. Your project files give the model your rules, your repeatable tasks, and your tools.

The model reasons; the harness transforms that reasoning into action by remembering context, retrieving information, validating results, executing changes, and orchestrating how all of these work together around the model.

One naming point prevents confusion. GitHub Copilot is a product, not a model, and Claude Code and Codex are clients too. Each puts a chassis around an engine.

Your client builds the chassis. You supply the one thing that it cannot know: the rules of your project.

Five pieces build your dashboard

Ask Copilot: “Add a field to the Customer table.” The answer looks right, but it can use the wrong object ID, skip a caption, or put logic in the wrong object. A better prompt fixes this once, not for the next task or the next developer.

Your dashboard fixes it for the whole team: your rules, one repeatable task, and one safe tool live in small files next to your code. Five pieces make a complete dashboard, but a rules file and one read-only tool are enough for day one.

Piece What it does AL example
Instructions file Rules the model always sees “Object range 50100-50149. Use PascalCase for object names.”
Skill One task the model runs on request “Write a test outline for this table extension.”
Custom agent A focused helper with its own instructions and its own tools An agent that only reviews AL code against your rules, and never edits a file
Prompt A saved question you reuse A prompt that checks one codeunit against your rules
mcp.json Connects one outside tool the model can call A read-only lookup on Microsoft Learn

Here is how one request moves through these layers:

1
Your dashboard
Rules, a task, one tool: the files you add.
2
The chassis
The harness builds the request, calls the tools, and returns the results.
3
The engine
The model reasons, and writes the answer or the AL code.
✓
You read it, test it, decide.

A read-only tool you can add today

Microsoft Learn publishes a free, read-only MCP server for AL and Business Central questions. Put this in .vscode/mcp.json, or the equivalent file for your client:

{
  "servers": {
    "microsoft-learn": {
      "type": "http",
      "url": "https://learn.microsoft.com/api/mcp"
    }
  }
}

Open Copilot Chat and ask: “Check this codeunit against my rules. Use Microsoft Learn if you need an AL fact.” That is a complete dashboard: one rules file, one tool. Add a skill, a custom agent, or a prompt only when you repeat the same request.

The same dashboard, packaged for other clients

Every client reads a dashboard in its own way. Claude Code reads CLAUDE.md, Codex reads AGENTS.md, and the GitHub Copilot CLI reads both. Each also has its own MCP config file.

Copying files by hand works for one repository, not for five. Two packages fix that:

  • A VS Code extension. It bundles your instructions, skills, and MCP servers into one install, for the VS Code client only.
  • An Agent Plugin. The Agent Plugins 1.0 standard packages skills and MCP servers once for Claude Code, Codex, and the Copilot CLI. A custom agent still needs its own file per client. Read Agent Plugins 1.0: One AL Plugin for Three Copilot Clients for the details.

Closing note

Your client delivers the engine and the chassis. You build the dashboard from five small pieces. Start with two, the rules file and one read-only tool, and add the rest when you need them.

← Back to AI