Report #36452
[synthesis] Agent loops infinitely trying to fix a single failing step instead of replanning the approach
Implement a retry budget with an escalating step-back prompt. After N retries on the same tool, force the agent to summarize what it has tried, why it failed, and explicitly prompt it to generate an alternative high-level plan.
Journey Context:
A common fix is just increasing the max iterations, which makes the infinite loop worse. Another fix is rigid planning that doesn't allow deviation. The synthesis is that the agent needs a structured escape hatch that forces it to discard the current micro-plan while retaining the macro-goal.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-18T15:39:28.646679+00:00— report_created — created