Build, run, and improve a grounded AI agent in 90 minutes with Microsoft Foundry. You'll pick a model, build TrailMate (a product expert for Contoso Outdoors), measure it, then climb the hill — improving one lever at a time with Model Router and Agent Optimizer, and promoting the version the evidence supports.
TrailMate — a prompt agent that answers questions about Contoso Outdoors gear using the product manuals (via a file-search tool), and refuses to invent anything the manuals don't say.
|
You ask:
TrailMate answers (grounded):
It cites the right manual for each product, and if you ask something the manuals don't cover (like a warranty), it says so instead of guessing. Getting it to do that reliably — and proving it — is the workshop. |
You'll evolve TrailMate through three measured versions, each changing one thing:
| Version | One change | Lesson |
|---|---|---|
| v1 · Basecamp | Frontier GPT + simple instructions + file search | Build a grounded agent; measure a baseline |
| v2 · Trailfinder | Model → Model Router | Cut cost/latency without losing quality? |
| v3 · Summit | Instructions → optimized (Agent Optimizer) | Improve grounding and honesty |
By the end of the core workshop you'll be able to:
- Right-size a model for a task by capability, cost, latency, and quality.
- Build a grounded prompt agent with a file-search tool over your own documents.
- Observe agent behavior with traces and run metrics (AgentOps).
- Measure a baseline with a rubric evaluator.
- "Hill climb" — improve one lever at a time (Model Router, then Agent Optimizer) and promote the version the evidence supports.
- A personal GitHub account (for Codespaces — a work/enterprise account won't work).
- A modern web browser.
- Azure access to Microsoft Foundry:
- Instructor-led (Skillable): provided for you — nothing to install.
- Self-guided: your own subscription with Foundry access.
- The dev container sets up Python 3.13, Azure CLI,
azd, and every notebook package for you.
Models used (Skillable pre-provisions all of them; the self-guided
provision.sh deploys the Azure-direct ones and prompts
for Claude):
| Deployment | Role |
|---|---|
gpt-5.4 |
Frontier reasoning; TrailMate v1 model |
gpt-5.4-mini |
Fast/cheap contrast in model selection |
model-router |
TrailMate v2 model lever |
MAI-Image-2.5-Pro |
Image task in model selection |
claude-sonnet-4-6, claude-haiku-4-5 |
Optional GPT-vs-Claude comparison |
Setup differs by track; the labs are identical.
- Instructor-led (Skillable) → instructions/skillable/README.md — infra is pre-provisioned; you generate
.envand go. Start here for the workshop. - Self-guided (bring your own Azure) — provisions infra with
scripts/provision.sh. Being finalized; follow the Skillable steps as reference for now.
After setup, open the core notebooks in order:
| # | Lab | Time | Notebook |
|---|---|---|---|
| 00 | Validate setup | 15 min | labs/core/00-validate-setup.ipynb |
| 01 | Select the right model | 30 min | labs/core/01-model-selection.ipynb |
| 02 | Build, evaluate & optimize the agent | 45 min | labs/core/02-agent-optimization.ipynb |
Have time left? The optional labs/more/ notebooks go deeper on specific capabilities — pick any that interest you, in any order:
| # | Lab | Time | Notebook |
|---|---|---|---|
| 01 | Multimodal chat | 20 min | labs/more/01-chat-multimodal.ipynb |
| 02 | Reasoning models | 20 min | labs/more/02-reasoning-models.ipynb |
| 03 | Image generation | 20 min | labs/more/03-image-generation.ipynb |
| 04 | Model Router | 20 min | labs/more/04-model-router.ipynb |
| 05 | Code with Claude | 20 min | labs/more/05-code-with-claude.ipynb |
foundry/agent-builder/
├── README.md you are here
├── assets/ banner, product images, attribution
├── instructions/
│ └── skillable/ Skillable setup (self-guided coming)
├── labs/
│ ├── core/ the 90-minute workshop — run in order
│ │ ├── 00-validate-setup.ipynb
│ │ ├── 01-model-selection.ipynb
│ │ ├── 02-agent-optimization.ipynb
│ │ └── assets/ screenshots for the core labs
│ └── more/ optional extra labs
│ ├── 01-chat-multimodal.ipynb
│ ├── 02-reasoning-models.ipynb
│ ├── 03-image-generation.ipynb
│ ├── 04-model-router.ipynb
│ ├── 05-code-with-claude.ipynb
│ └── assets/ screenshots for the more labs
├── scripts/ provision.sh · setenv.sh · sample.env
└── src/
├── data/ manuals · evaluation-cases.jsonl · evaluators/
└── agent/ build_agent.py · switch_to_router.py · instructions · VERSIONS.md
- Microsoft Foundry documentation
- Foundry models
- File search tool
- Model Router
- Agent Optimizer overview
- Optimize a prompt agent (quickstart)
- Evaluation in Microsoft Foundry
Part of Model Mastery · Browse Foundry workshops