How I Built an AI Agent That Investigates Cloud Bill Spikes (Architecture Inside)

I built Ghost-hunter, an AI agent that investigates cloud bill spikes the way a senior on-call engineer would: one hypothesis at a time, read-only, every command validated before it runs. The core design choice is that it's two agents, not one — a lead detective (Claude Opus) that forms theories and never touches your cloud, and an evidence technician (Claude Sonnet) that drafts commands which then pass a seven-gate validator before anything executes. This post is about why that split matters, h