Microsoft presents: Agentic development with AL developer tools and ALMCP deep dive

Overview

The session covers how coding agents work with AL tools. It explains Agent, Ask and Plan modes, custom agents defined in markdown files, skills loaded on demand, and the agentic loop where an agent picks tools, reads results and continues. Demos show building, publishing, debugging and troubleshooting, including access to stack frames, variables and database statistics during a debug session.

The second half covers the ALMCP server and AL CLI tool (.NET tools installable via NuGet), registering ALMCP in GitHub Copilot CLI, and the Language Server Protocol for AL. It then covers BCbench, which uses hidden tests from real commits, and a BC fix workflow built from sub-skills with a RAL loop, sub-agent context isolation, a session retrospective skill, and notes on performance profiling and pull request comments.

Key points

• Custom agents are markdown files with front matter (name, description, tools, skills, knowledge, system prompt). Skills load on demand and do not use context window space up front.
• Troubleshooting tools come from a dedicated MCP server and are only available during a debugging session. They expose source, stack frames, variables, and SQL statements and database statistics.
• ALMCP server and AL CLI are .NET tools installed via NuGet, with a pre-release flag for the latest features. In GitHub Copilot the agent may call ALC directly, so instructions may be needed to redirect it to ALMCP.
• LSP gives agents semantic endpoints such as go to definition, find references and rename. In the demo, grep missed about 80% of references to a number field on the Customer table. LSP needs a package cache path configured.
• BCbench is based on SWE bench and real commits that contain both a fix and a test. It runs the hidden tests against agent fixes and averages five runs because of LLM nondeterminism. Baseline was 67%, and adding tools raised it to 71%.
• The BC fix workflow uses a RAL loop with success criteria and a maximum number of attempts. The baseline skill creates a failing test, and the implement skill runs as a sub-agent on Claude Sonnet 4.6 to limit context rot and cost.
• Pull request comments by an agent still need human approval. Profiler reading by agents is not yet in the workflows, and a loop for pipeline gate failures would need to be added.

Features

A status with a roadmap link comes from the Microsoft 365 roadmap feature this part of the video covers (matched by Haiku; links Opus dropped are not used); other statuses need a status word in the video itself.

AL objects mentioned

As heard in the captions. A name that matches one object page by exact type and name links to it; the others stay as named.

• page “page (shown in demo for JSON parsing)” at 7:12
• codeunit “process from JSON method” at 7:23
• other “process entry method” at 7:23
• other “trim method” at 13:56
• other “lookup method” at 10:27
• table “customer table” at 21:26
• other “AL command” at 22:12
• other “ALC” at 15:46
• other “BCLdev session” at 29:44
• other “BC fix baseline” at 31:27
• other “BC fix implement” at 30:57
• other “BC fix supplement” at 31:27
• other “BC test implement” at 30:57

Not found in BC28-30: page “page (shown in demo for JSON parsing)”, codeunit “process from JSON method”, table “customer table”.

Quotes

• 1:06 “with the agent mode, you can have this multi-step operation where you can take it from a small simple operation …”
• 1:52 “we can also make our own custom agents. The the beauty of this is that it’s all just a markdown file”
• 2:43 “it doesn’t take space in your context window. All it is from uh for loading it for the agent to begin with is a”
• 3:30 “what has been truly a gamecher with agents in comparison to the chat modes that we’ve seen in the past is their ability to”
• 3:50 “it will discover them and as soon as you give them a task it will reason over that task what it has to do”

Disclaimers in the video

• 15:17 preview: you can also specify the pre-release flag so that you get access to the latest uh bits that we’re shipping
• 32:29 preview: I’m going to skip the planning mode completely. It was shown in the keynote very fast.
• 34:30 other: It’s probably going to run on the hik model as well. I haven’t tested this one. That’s exactly why we need BCbench for this.
• 38:00 other: the part that I will um I’m going to skip I think for the pull request I have the small …
• 42:48 not-in-this-release: We are not doing this right away. We don’t have a workflow for this one but we have it in mind
• 44:39 coming-later: we need probably to add like another loop to make sure that agent can help with the failures in the gates

Presenters (as heard): Thomas, Quinton, Arthur.