Skip to content

Stage 7 · Foundry

The pro-code frontier. When a low-code agent can't reach far enough — custom models, autonomous orchestration, evaluation, and MCP tools at scale — you cross into Microsoft Foundry.

The shift: from "I build agents in a designer" to "I engineer agents as software." Most people never need this stage, and that's by design — Stages 1–5 solve the overwhelming majority of real work. But when an agent has to run pro-code logic, call custom or fine-tuned models, or operate as part of an engineered system, Foundry is where it belongs.

A different kind of stage — for developers

Foundry is a pro-code, developer-owned platform, not a low-code designer. It lives in Azure, not the Microsoft 365 admin center, and it assumes engineering skills (SDKs, CI/CD, Azure resources). Treat this stage as a signpost to the frontier rather than a click-by-click walkthrough — and verify everything against the official Microsoft docs, which move fast.

When you're ready for this stage

You've outgrown low-code. A Copilot Studio agent can't express what you need: you require a custom or fine-tuned model, autonomous multi-agent orchestration, rigorous evaluation and monitoring, or MCP tools and integrations at engineering scale. You have developers, an Azure subscription, and a deployment pipeline — and the problem justifies building like software.

Going the wrong way?

Foundry is the easiest place to over-engineer. If you're writing glue code for a single linear task or rebuilding a connector Studio already ships, you've overshot — drop back to Studio.

What you need

  • An Azure subscription and a Microsoft Foundry project to build in.
  • Developer skills — SDKs (Python/.NET/JS), source control, and CI/CD.
  • An engineering owner for the Azure resources, identity, and networking behind the agent.
  • A clear reason you couldn't solve it in Stage 6 · Copilot Studio. If you could, start there.

Start here — if you only do three things

Prove the plumbing with a first SDK agent, re-platform a real agent that hit a ceiling, then put an evaluation gate around it before it ships.

All walkthroughs

Unlike earlier stages, these are developer walkthroughs — representative SDK/CLI patterns and the decisions behind them, not click-by-click product tours. Foundry moves fast; each page links the authoritative Microsoft docs to verify against.

Build your first pro-code agent

Stand up the agent, re-platform a proven one, and give it real tools.

Orchestrate, evaluate & govern

Move from one agent to a governed system you can trust in production.

Operate it in production

Go past building to operating — swap models safely, keep it affordable, and prove it's hard to break.

Build something deployable

These walkthroughs pair with copy-and-adapt Foundry solution templates — a representative code scaffold, an evaluation gate, and a deployment checklist. Reach for one only when a Studio agent genuinely can't carry the job.

  • Pro-Code Grounded Q&A Agent

    Custom retrieval, an evaluation gate, and code-owned identity for grounded Q&A that outgrew Studio.

    ⏱ 1–2 days to build ⚙️ Adapts to: any grounded Q&A workload

  • Multi-Agent Workflow Orchestrator

    Automate a multi-step process across specialist agents with deterministic control and a workflow-level eval gate.

    ⏱ 2–4 days to build ⚙️ Adapts to: research → draft → review, intake → enrich → route

  • High-Volume Document Processing Agent

    Extract structured data from documents at scale, with schema validation and a precision/recall gate.

    ⏱ 2–4 days to build ⚙️ Adapts to: invoice/PO extraction, contract clauses, claims intake

  • Customer-Facing Support Agent

    An external, action-taking support agent with strict guardrails, content safety, and continuous evaluation.

    ⏱ 3–5 days to build ⚙️ Adapts to: order/account support, entitlement checks, ticketing

  • NL-to-SQL Analytics Agent

    "Ask your data warehouse" self-service analytics with a SQL-safety layer and a correctness gate.

    ⏱ 2–4 days to build ⚙️ Adapts to: self-service BI, ops/finance/product analytics

  • Browser-Using (Computer-Use) Agent

    Automate web apps that have no API — sandboxed, allow-listed, and human-confirmed on irreversible actions.

    ⏱ 3–5 days to build ⚙️ Adapts to: legacy web apps, form-filling, portal data-gathering

  • Voice Agent / Contact-Center IVR

    A real-time phone agent — speech-to-text, grounded conversation, text-to-speech, clean human handoff.

    ⏱ 3–5 days to build ⚙️ Adapts to: support lines, booking lines, helpdesk hotlines

  • Code-Review / PR-Triage Agent

    Advisory, private-repo PR review and reviewer routing — comments only, never merges.

    ⏱ 2–4 days to build ⚙️ Adapts to: PR triage, review assistance, reviewer routing

  • Multi-Modal Document Classification Agent

    Sort a high-volume document stream by vision + text, with confidence-based routing and a precision/recall gate.

    ⏱ 2–4 days to build ⚙️ Adapts to: intake routing, claims triage, content moderation

Borrow, don't build

Foundry is a deep, fast-moving developer platform — there's no value in re-documenting it here. Lean on Microsoft's own authoritative material, which stays current with the product.

The full curated set for this stage lives in Resources → Stage 7.

Security at this stage

Foundry agents run on Azure with Entra ID authentication and Azure RBAC, and the standard setup keeps agent data in your own single-tenant Azure resources (Storage, Cosmos DB, AI Search, Key Vault). Microsoft Purview classifies them as Enterprise AI apps, and Agent 365 can govern them alongside your low-code agents. See the full picture in Security & Governance.

You've reached the frontier

There's no Stage 7. From Chat to Foundry, the ramp has taken you from a daily habit to engineered, governed agents — pick the lowest stage that solves each problem, and only climb when the work demands it. Keep a backlog of agents worth building.

Back to the full Use-Case Catalog →

Was this page helpful?

Found an issue or want to suggest an improvement? Open a GitHub issue  ·  Star on GitHub