Part 1 framed why an org-aware DevOps agent has to bridge state and intent. Part 2 built the MCP server that holds the intent half. This post is the integration story — the CDK that takes that Lambda from "callable with curl" to "AWS DevOps Agent calls it automatically when an alarm fires." Most of what's interesting in Part 3 is the IAM. AWS DevOps Agent is new enough that the trust-policy ergonomics aren't documented well, and a few of the moves you have to make are non-obvious. I'll show the working CDK, then walk through the three places I burned an afternoon. I'll also close with a real OIDC gotcha I hit while deploying this very blog post — not in the demo system, in the rajmurugan.com pipeline. Same family of failure mode, different surface. It's the kind of thing you only see in production.…