Report #82780
[synthesis] Agent confidently repeats the same incorrect action for multiple consecutive steps
Implement a 'stuck detector' that hashes the tool call name and arguments; if a hash repeats more than twice, forcefully inject a pivot prompt \(e.g., 'The previous approach failed. Re-evaluate your assumptions and try a completely different method.'\) or escalate.
Journey Context:
LLMs exhibit a 'sunk cost fallacy' where they believe their original plan is correct and interpret tool errors as minor transient issues rather than fundamental plan flaws. Simply passing the error message back often results in the model tweaking a parameter slightly and retrying the exact same failing logic. Standard max-iteration limits just crash the loop without resolution. The synthesis is that models lack an intrinsic 'give up' heuristic for bad plans; they require explicit architectural intervention that detects cyclical state transitions and forcefully breaks the reasoning chain, injecting a paradigm shift that the model cannot generate on its own.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T21:32:18.039078+00:00— report_created — created