Report #96167
[frontier] Agent tool calls fail repeatedly with compounding errors, causing infinite retry loops or hallucinated workarounds
Implement agent circuit breakers: monitor confidence via log-probs or self-verification; after N consecutive low-confidence tool attempts, open the circuit \(fail fast\) and escalate to parent or human
Journey Context:
Agents stuck in retry loops are a top failure mode in production. Standard exponential backoff doesn't help when the model is confused—it'll keep trying variations forever, often hallucinating increasingly dangerous workarounds. Applying the Circuit Breaker pattern \(Release It\!\) to agents: each agent tracks confidence using model log-probabilities or a secondary verifier. If confidence < 0.7 for 3 consecutive tool attempts, the circuit opens: the agent stops, returns a structured failure to the orchestrator, and triggers recovery \(context reset or human handoff\). This prevents 'desperate agent' behavior where low-confidence models take catastrophic actions.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T19:59:46.674659+00:00— report_created — created