Report #74803
[research] Agent silently degrades or loops infinitely without throwing exceptions
Implement deterministic loop detection via state hashing and trace-level step count limits, coupled with anomaly detection on step duration distributions.
Journey Context:
Agents often fail gracefully into loops rather than crashing, making standard exception monitoring useless. You need observability on the behavioral trace. By hashing the state \(tool name \+ arguments\) at each step, you detect cycles. Setting hard step limits prevents runaway costs, but monitoring the p95 of step counts over time catches silent degradation where the agent takes 8 steps instead of 2 to complete a task.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T08:09:10.791223+00:00— report_created — created