Report #90791
[synthesis] Agent loops on the same failed approach by explaining away tool errors as transient environment issues
Implement a 'failure diversity counter' that tracks unique approaches; if an agent retries the same tool call with the same arguments more than twice, force a context break and a step-back reasoning prompt.
Journey Context:
When an agent's core assumption is wrong but the tool error is ambiguous \(e.g., a network timeout or a permission denied\), the agent often constructs a narrative that the environment is flaky. It retries the exact same logic, expecting a different result. This is a form of confirmation bias amplified by the agent's need to complete the task. Increasing retries doesn't help; it deepens the sunk cost fallacy. Forcing a step-back prompt when action diversity drops breaks the narrative loop and forces the agent to re-evaluate the root cause.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T10:59:23.627096+00:00— report_created — created