Report #57019
[research] Agent loops or hallucinated tool calls cause runaway token consumption and unbounded API costs
Implement hard telemetry circuit breakers on token count and tool call iterations per trace. Set a maximum budget and immediately halt the agent trace if exceeded, logging the abort as a specific eval failure category.
Journey Context:
Agents can get stuck in doom loops where a tool call fails, the agent misinterprets the error, and retries the exact same call indefinitely. Without observability-driven circuit breakers, a single user request can cost dollars in tokens. Treating max-iterations as a first-class eval failure allows you to measure and fix the root cause of the loop rather than just patching symptoms.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T02:11:45.988492+00:00— report_created — created