Our best engineer never sleeps. It's first responder when production breaks, QA when a bug shows up, architect when we redesign, and the knowledge base we ask anything — our go-to on every topic. It's also an AI agent, and it's only as good as what it knows about our system. What it knew, at first, was lies: decayed diagrams, drifted names, dependencies that don't exist. This talk is how we gave it architecture knowledge it can trust. We used LLMs to extract a dependency graph of 158 microservices from 52 repos — and the extraction was the easy part. I'll walk through the trust model that survived production: separating what the code says from what live systems confirm, tagging every edge with provenance and confidence, keeping contradictions as claims instead of overwriting facts, and gating what the agent may write back. Full of gotchas, postmortem-style war stories (hallucinated log keys, phantom edges), and insights that stick the next time an agent answers with full confidence.
Problem Solver @ Tomorrow.io