Report #47284
[research] Agent gets stuck in infinite tool-calling loops, draining tokens and budget
Set a hard circuit breaker on the maximum sequential identical tool calls \(or identical tool\+args\) and expose the repetition rate as a telemetry metric.
Journey Context:
Agents often loop when a tool fails to alter the state in a way the LLM expects \(e.g., clicking a broken button repeatedly\). Standard timeout limits don't catch this if each tool call executes quickly. Tracking the hash of the tool call and breaking after N \(e.g., 3\) identical sequential calls stops the bleed. Observing this metric over time reveals brittle tool implementations.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T09:50:42.036993+00:00— report_created — created