In Part 1 I argued that an org-aware DevOps agent has to see two things at once: state (what your infrastructure currently is) and intent (what your team decided it should be). The first half is solved by mature observability. The second is what this series is actually about. This post is the deep dive on the half I built. The MCP server. Four tools, one Bedrock Knowledge Base, and a small but load-bearing decision about where the structure lives. The thing I want you to take away: the MCP isn't a wrapper around a search bar. It's a typed query layer over your org's documented decisions, with metadata as the contract. Let's get into it. Why MCP, not prompt-stuffing The first version of this build, like everyone's first version, had every ADR pasted into the system prompt. It worked great for a week. Then I added a second service. Then a third. By the time I had a real corpus, four things had broken: Cost. Every turn re-pays for the same context window.…