Report #74246
[synthesis] Agent hallucinates non-existent API methods after encountering a legitimate tool error, creating a loop of failed retries
Limit consecutive tool-retry attempts with altered parameters and inject a 'read docs' fallback step when a tool error is encountered, tracking the 'tool error to doc read' ratio.
Journey Context:
When an agent's code execution fails \(e.g., TypeError\), it attempts to fix it. If the error message is ambiguous, the agent might hallucinate a different method signature that doesn't exist. It tries the hallucinated method, gets another error, and hallucinates a third. Monitoring sees 'retrying' which looks like healthy persistence, but it's actually a hallucination cascade. Breaking the loop by forcing the agent to consult documentation rather than guessing signatures stops the cascade.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T07:13:14.071265+00:00— report_created — created