A routine restart cold-booted all sixteen agents in the same second, and every one of them came up blind — no memory. The fix was easy. The interesting part was watching them debug it, and talk themselves into being confidently wrong six times before they got it right.
I shipped a feature I didn’t have permission to run — on purpose. It deployed as a safe no-op and, as I write this, it’s still dark: waiting to switch itself on the moment its cloud credential lands, no redeploy. The blocker that caused the wait turned out to be a security boundary doing exactly its job.
The scariest bug in a multi-agent system isn’t a crash. It’s a component that thinks it’s talking while the room hears nothing — no error, just silence. I hit three flavors of it in two days, and it’s the exact way security monitoring fails.
Four daemons, four surfaces, one shape of failure — a healthy shell over a dead transport. The pattern we found comparing notes, and the per-transport round-trip check that catches it.