deploying-scalable-agents

Deploying Scalable Agents with Microsoft Foundry

Companion skill for Lesson 16 – Deploying Scalable Agents. Use it to help a learner move an agent from prototype to a scalable, observable production deployment. Ground every recommendation in the lesson content and the runnable notebook; do not invent Foundry APIs.

Triggers

Activate this skill when a learner wants to:

• Deploy an agent to Microsoft Foundry as a hosted agent and make it versioned/observable.
• Choose between client-hosted, hosted-agent, and agent-workflow deployment patterns.
• Add model routing, response caching, or bounded concurrency to control latency and cost.
• Add an evaluation gate so a bad agent version cannot ship.
• Add a human-in-the-loop approval step for high-risk actions.
• Instrument an agent with OpenTelemetry tracing for production observability.
• Smoke-test a deployed agent as a fast post-deploy gate.

Core mental model

A production agent is mostly the operational skeleton around the model (~80%), not the model itself. Map every recommendation to one of these concerns:

Deployment patterns (pick one, or combine)

1. Client-hosted — the reasoning loop runs in your process. Max control; you own scaling/state.
2. Hosted agent (Foundry Agent Service) — Foundry hosts the loop, stores threads, enforces RBAC/content safety, shows the agent in the portal. Less control, far less operational surface.
3. Agent workflow — multiple agents/tools composed into a graph with branching, approval nodes, and durable checkpoints.

Lifecycle (the loop that ships an agent)

create → version → evaluate (gate) → deploy hosted → observe online → collect failures → repeat. Offline evaluation is a gate, not an afterthought — a version does not ship unless it clears the threshold. Online observability feeds real failures back into the offline test set.

Scaling and cost levers (in priority order)

1. Right-size the model — use the smallest model that passes the evaluation gate.
2. Route by complexity — small/fast model for simple requests, large model for real reasoning (DIY classifier or Foundry Model Router).
3. Cache — serve near-duplicate requests without a model call.
4. Stateless design + bounded concurrency — externalise state; retry with backoff.

Key patterns to reproduce

Point the learner at these from the notebook 16-python-agent-framework.ipynb:

• Request handler: cache → route by complexity → trace span → run → cache.
• Evaluation gate: score an offline test set; return pass_rate >= threshold and only deploy if true.
• Human approval: @tool(approval_mode="always_require") for actions like large refunds.
• Tracing: wrap each request in tracer.start_as_current_span(...) and set attributes like routed.model, customer.id.

Smoke-testing a deployed agent

After deploy, verify the endpoint actually answers (a green deploy can still be silent). Use the AI Smoke Test action via .github/workflows/smoke-test.yml with the catalog in tests/. The runner POSTs each prompt to POST {project_endpoint}/agents/{agent_name}/endpoint/protocols/openai/responses and asserts on the reply text. The identity needs the Azure AI User role at Foundry project scope; the token audience must be https://ai.azure.com/.

Layer the gates: smoke test (reachable/responding, every deploy) → offline evaluation (good enough to ship, before promotion) → online evaluation (how is it doing in the wild, continuous).

Enterprise controls

• RBAC: give each hosted agent a managed identity with least privilege.
• MCP in production: treat every MCP server as an untrusted boundary — pin the version, scope its identity, validate outputs, rate-limit, never expose secrets.

Guardrails for the assistant

• Prefer the canonical FoundryChatClient(...) + provider.as_agent(...) pattern used across the course.
• Do not promise live-Azure results you have not verified; recommend the smoke-test workflow to confirm a deployment.
• Keep evaluation and cost advice tied together: evaluation sets the quality floor, routing/caching keep cost near that floor.