Self-Service 'Deploy a New API to APIM' via Multi-Agent Platform

🛠 Capstone lab from the AI Platform Engineering on Azure course. Doubles as the Azure-variant slot for the AI Platform Engineering Bootcamp Week 8 multi-agent capstone. Course landing page: https://academy.tekanaid.com/courses/ai-platform-engineering-azure Bootcamp landing page: https://academy.tekanaid.com/bootcamps/ai-platform-engineering-bootcamp

🟡 Beta lab. Hands-on instructions, check scripts, and solve scripts are in place.

This is the closing capstone for AI Platform Engineering on Azure. It assumes you have already taken Lab 1 (single agent + Azure MCP) and Lab 2 (custom APIM MCP server in C#). The lab starts by re-provisioning the Lab 1 + Lab 2 foundation on a fresh VM (Azure OpenAI + APIM Consumption tier + 2-node AKS, all pinned), then layers the multi-agent supervisor on top.

You will deploy a minimal version of Lab 1's Azure MCP agent and Lab 2's APIM MCP server side by side on AKS, add the official `ghcr.io/github/github-mcp-server:v1.0.3` wired to a local GitHub fixture (no real GitHub workspace required), and build a Microsoft Agent Framework Workflows supervisor with an explicit three-state graph: plan → approve → apply. Each state has a single specialist that calls one MCP server (APIM, GitHub, or Azure). The supervisor runs in dry-run mode by default — it prints the planned APIM mutation and the planned GitHub commit, then pauses for an approval message that lands in a Slack-shaped Flask fixture on port 9008.

The end-to-end flow exercises the full topology: you type "publish a new orders API with rate limit 100/min", the supervisor's plan state asks the APIM MCP what is already there and proposes the missing API, the approve state pauses for a Slack-fixture approval, and the apply state commits the APIM declarative manifest via the GitHub MCP and then creates the API revision via the APIM MCP. The lab closes with a guided concept walkthrough of Azure Front Door in front of APIM — diagrams, sample `az afd` commands you can run after the lab if curious, no student-side provisioning.

By the end you will have shipped a working reference architecture for agent-driven self-service infrastructure with the safety properties production demands: scoped per-agent tool surfaces, explicit state machine for the workflow, dry-run plus human approval gate before any mutation, and GitOps as the source of truth for the resulting APIM manifest.