We found a hole in our own egress story by explaining it out loud. The fix, the probe we ran on a live cluster, and what is still open.
We replaced a wall of claims with an interactive anatomy of a session. Click a part to see what it holds, or attack it and watch where it stops.
A coding agent reads untrusted text for a living. Ours runs with permissions bypassed and holds no credentials. The reasoning, and what's unsolved.
The second agent took an afternoon. The two bugs it surfaced were nowhere near the model.
Job tokens die with their jobs and deploy keys sprawl, so our agents stopped pushing entirely.
Kubernetes already replaces dead pods. Why were we running a manager, with a standing secrets-reader Role, to do it again?