A gaming-transaction pipeline into Kafka broke, on and off, for eight days — and looked like a client bug the whole time. What finally cracked it was changing how we investigated: five Claude Code sessions running in parallel, each owning one slice of evidence, coordinated by an agent that wasn't allowed to look at any data itself. A real production incident, the wrong turns along the way, and a practical playbook for running AI agents inside a live incident — including where they genuinely help and where they don't.
