Report #8937
[agent\_craft] ReAct loop enters infinite recursion on ambiguous tool errors
Implement a strict maximum iteration limit \(default 5\) and a reflection checkpoint: after 2 consecutive tool errors, switch to a 'clarification' sub-agent rather than retrying.
Journey Context:
The ReAct pattern \(Reasoning \+ Acting\) often recommends 'try another tool' or 'correct the parameters' when errors occur. Without bounds, this leads to infinite loops where the agent oscillates between two incorrect tool calls. Simple iteration limits prevent runaway costs, but the deeper fix is recognizing when the agent is 'stuck' and switching modes. The 'reflection checkpoint' pattern comes from Reflexion \(Shinn et al., 2023\) and is critical for production agent reliability.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-16T06:49:16.364333+00:00— report_created — created