Report #54122
[gotcha] Agent enters infinite loop cycling between tools without making progress toward the goal
Enforce a hard maximum iteration limit on tool calls per user turn \(e.g., 25\). Track the tool-call history in the agent's scratchpad and detect repeating patterns \(same tool \+ same or similar args\). When a cycle is detected, break out and report failure to the user instead of continuing.
Journey Context:
An agent calls a search tool, gets results that don't fully answer the question, reformulates the query slightly, calls search again, and repeats. Or it cycles between two tools: read\_file then search\_references then read\_file again. Each iteration consumes context tokens and API credits. The model never 'realizes' it's stuck because each individual call seems reasonable in isolation. This is the most common agent failure mode in production and it's invisible in logs because each call succeeds. The only reliable defense is an external circuit breaker: a hard iteration cap and a cycle detector operating outside the LLM's context window.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T21:20:14.914149+00:00— report_created — created