Report #8811
[research] Agent gets stuck in an infinite loop of tool calls or self-corrections, draining token budgets
Implement a circuit breaker based on a maximum depth of consecutive tool calls without a change in state, and alert on the repeated tool call telemetry metric.
Journey Context:
Agents often loop \(e.g., Tool A fails, Agent retries Tool A with the same args\). Standard timeout limits are too coarse; an agent can burn 100k tokens in 60 seconds. The observability fix is tracking state transitions. If the agent invokes the same tool with the same arguments more than N times, or if the trace shows more than M consecutive tool calls without emitting a final answer, trigger a circuit breaker and log a loop detected event.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-16T06:36:14.004887+00:00— report_created — created