Report #35662
[synthesis] Agent confidently repeats the same incorrect action for multiple consecutive steps
Implement a state-based retry limit that detects identical or semantically similar tool calls and forces a pivot by injecting a system message that explicitly challenges the current approach.
Journey Context:
When an agent makes a bad assumption \(e.g., wrong file path\), it often receives an error. Instead of re-evaluating the assumption, it tries slightly different arguments for the same tool, digging a deeper hole. This is the LLM equivalent of the sunk cost fallacy. Simply limiting retries isn't enough; the agent needs a circuit breaker that alters its reasoning trajectory, forcing it to step back and verify the premise.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-18T14:20:06.981737+00:00— report_created — created