Everyone teaches building agents. Almost nobody teaches running ai agents in production — the operational discipline that separates a demo from a fleet that survives contact with reality.
Why AI agents fail in production rarely makes it into the guides — they say 'log every decision path' and stop there. Here are 7 real failure modes from running 325 agents, and the exact signal that catches each one.
Your monitoring probes return green. Your deployment is actually blocked. Here's the failure mode Vercel's bot protection introduces and the pattern that catches it before production breaks.
A trading bot ran dead for seven hours before I noticed. That failure taught me that 'process alive' and 'system healthy' are not the same sentence. Here's how I built infrastructure that heals itself, documents itself, and gets smarter every time it breaks.
Three hundred and forty tests. Ninety-five percent coverage. Five docs, a runbook, a watchdog, and a launchd plist. In three weeks it had never placed a real trade. This is what we learned when we finally pulled the data instead of writing another fix.