Labwrights

Changelog

What changed in this package, newest first. Before every release, each lab is executed again on live Azure.

[2027 Edition — Early Access, update 1] — 2026-10-27

A documentation-only update; no lab content changed.

  • START-HERE.md, section 2, now says plainly that the labs run on your own Azure subscription and bill you directly. It also shows how to check which subscription the labs will use, and change it, if you have more than one.
  • Azure costs: START-HERE now say that your costs depend on your use, and that they're entirely yours: Azure bills you directly.

[2027 Edition — Early Access]

First release.

  • 11 labs that take a working app from first run to a production AI agent on Azure: tool calling, the agent loop with its limits, approval for every action, Microsoft Agent Framework, your own MCP server secured with Entra ID, a Foundry agent sharing it, an evaluation gate for CI, tracing and alerts, and layered defences against injected instructions.
  • Every command and expected output was captured from a real run on live Azure in October 2026.
  • Written to be read by newcomers: each lab explains its ideas in plain paragraphs with an analogy, and lists its new words up front. It also walks through all the code you write, in short commented pieces. GLOSSARY.md collects every term in one place.
  • The AI Engineer's Handbook (handbook/): 66 topics in 8 chapters, on LLMs, agents and tools, MCP, agent evaluation, production, safety and security, and Azure. Each is tagged beginner, intermediate, or advanced, with links for further reading. Checked against official sources in October 2026.
  • A capstone with an autograder (capstone/): give your agent an approval-gated rollback_deploy tool to a written contract, graded offline by scripts/grade-capstone.sh, including a recorded real model run replayed through your agent. No solution ships; the grader is the answer key.
  • Extras (extras/): a 58-question interview pack tied to the labs, a troubleshooting index of every error the labs can produce, and the glossary as a flashcard deck (scripts/make-flashcards.py regenerates it).
  • Reference solutions for the end of every lab (solutions/), for the backend and the MCP server, plus scripts/catch-up.sh to put your code at any of them, scripts/ask-agent.py to watch a run from a terminal, and scripts/warning-signs.sh to run the monitoring queries.
  • Pinned dependencies: the backend and the MCP server install through their constraints.txt files, the frontend through app/frontend/package-lock.json, so the versions you get are the versions that were verified. Key versions: openai 3.39.0, agent-framework-core 1.23.0, azure-ai-projects 2.7.0 (Agent Framework 1.23 needs a version below 2.8), mcp 2.3.0, azure-ai-evaluation 1.18.10, azure-monitor-opentelemetry 1.8.11.
  • Model version: the agent runs on gpt-5.4-mini (2026-03-17). Azure's catalog lists its retirement as 2027-09-21; lab 2 shows how to check and switch models.
  • Known limitations:
    • Lab times are estimates.
    • The CI evaluation-gate workflow (.github/workflows/agent-eval-gate.yml) ships with a manual trigger, ready to enable in your own copy; its one-time setup is in lab 9 and in the file's header comments. It was run in GitHub Actions in October 2026 (8 tasks passed).
    • The MCP server's rate limit counts in memory, which is correct for the one replica the labs run. With more replicas, lab 11 explains the options.
    • Foundry Agent Service's own guardrails for tool results were in preview in October 2026; lab 11 builds its defences in your code instead.