Add Lab D: observe, evaluate, and secure your agents

[ivorb]

Adds Lab D — Observe, evaluate, and secure your agents, closing the biggest gap in the catalogue: there is currently no observability, evaluation or security-testing content anywhere in the repo.

Tasks

	Task	Level	Time
Core	Trace your agent — instrument with OpenTelemetry and read traces in Foundry Monitor	L300	~25 min
Core	Evaluate answer quality — score a grounded agent against ground truth with the built-in evaluators	L300	~35 min
Optional	Red team your agent — attack strategies and custom seed prompts via azure-ai-evaluation[redteam]	L400	~35 min

Plus a Getting started page. Core path is 60 minutes.

What's included

• 5 instruction pages following the A/B/C structure, with the task table generated by tools/generate_lab_blocks.py
• 14 starter files with fill-in-the-blank placeholders, and 15 completed solution files
• A 10-row ground-truth JSONL for evaluation, and custom attack objectives for red teaming
• setup/check_env.py and setup/bootstrap_agent.py, so each task can be started on its own
• Lab B's knowledge documents copied in, so Lab D stands alone
• One entry in Labfiles/_shared/manifest.yml; the four infra files are generated by sync.py

Not validated end to end

Nothing in this lab has been run. No trace was exported, no evaluation scored, no scan executed — there was no Azure subscription available while writing it.

What was checked: every API call against the shipped SDK source and current Learn docs, Tier 0, sync.py --check, and the Tier 1 contract (imports resolve against azure-ai-projects==2.3.0 and azure-ai-evaluation==1.18.3, pip check clean).

It needs a walkthrough against a real subscription before it can be considered tested. The frontmatter carries status: 'draft' accordingly.

Choices worth a reviewer's eye

1. Application Insights is connected in the portal, not provisioned in Bicep. Adding it to the shared template would have given Labs A, B and C a resource they never use. D0 has the learner connect it via Agents > Traces > Connect, which is where they then read traces.
2. Task 3 attacks through a callback into the deployed agent, not a model config, so it exercises the same grounded agent that Task 2 evaluated.
3. agent_target.py is provided complete, so Task 2 is about evaluation rather than re-teaching agent calls the earlier labs already cover.

Region and preview constraints for the Red Teaming Agent are surfaced at project-creation time in D0 and repeated in D3.

No CI/CD content: that was cut deliberately as the wrong persona for a developer learning path.