Nine real failure modes from running an agentic publishing pipeline daily: symptom, root cause, and the guard now in place for each one.
I use Claude Code every day, in my own editor, on personal projects and client work alike. Sustained daily use produces something that launch-day write-ups never have: a failure log. Nine distinct failure modes, each one paid for in real debugging hours, each one now behind a guard.
They are worth writing down because they are not bugs in one tool. They are structural patterns of agentic systems, and any team putting an agent into production will meet most of them. To keep the account measured rather than anecdotal, I audited my own session logs: eleven production sessions, 9,135 tool calls, and every result the tooling flagged as an error. That baseline produces the most useful finding in this article, which is that the failures you can count are the cheap ones. ## What actually breaks when you run an AI agent in production?