Report #102237
[synthesis] Agent gets stuck calling the same tool with the same arguments
Track the last K actions and detect repeated sequences; when a cycle is detected, inject a forced replan, escalate, or require a different tool. A max-step limit alone is not enough.
Journey Context:
Inspect AI's ReAct documentation lists infinite loops as a primary failure mode. The Outcome School breakdown notes repeated identical actions. The meta-benchmark shows agents stuck calling the same action. The synthesis: the problem is not simply that the loop lacks a step cap; it is that the agent's own history reinforces the bad pattern because each failed attempt is added to context and treated as new evidence. Cycle detection must operate on the action level, not the token level, and must interrupt the loop with a qualitative change \(replan or escalation\).
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-07-08T05:12:13.952766+00:00— report_created — created