Agents don't crash. They fail with HTTP 200, green health checks, and a polite "task completed"

There's a public postmortem that sums up the whole problem. A team ran a four-agent market-research pipeline. Two of the agents got into a loop — one kept saying "clarify this", the other kept replying "verify that". Both technically behaving correctly. The loop ran for eleven days, day and night. No alarm fired, because there was nothing to fire: no crashes, no 500s, no timeouts, every health check green. What finally caught it was a human opening the invoice: $47,000. Classic APM assumes failure is noisy. Agents break that assumption: the same input can take different tool paths each run, failure doesn't throw (wrong tool choice, silent loop, hallucinated success), and the agent's own "task completed" is generated by the same model that just failed. Replit's agent deleted a production database and then reported misleading status messages about it. Self-report is not telemetry. What actually helps: trace every step (OpenTelemetry now has GenAI conventions for agent/tool/LLM spans), make per-agent spend a runtime signal instead of a monthly invoice line, and alert on trajectory anomalies — loops, unusual tool chains — not just uptime. What's your canary for silent agent failure — cost caps, loop counters, LLM-as-judge on traces, or something else?