Report #35785
[architecture] Agents enter infinite handoff loops because neither agent has the context or authority to resolve the task
Implement a maximum handoff depth counter and an escalation protocol; if an agent receives a task it cannot resolve and must hand it off, the handoff must include a strictly increasing depth field that forces termination or orchestrator intervention at a threshold.
Journey Context:
LLMs are eager to please and often lack the global view of the workflow. When an agent realizes it cannot do a task, it hands it off. If the recipient also cannot do it, it hands it back. This creates a tight loop of API calls that burns tokens rapidly. Adding a depth or hop\_count integer to the message metadata allows the orchestrator to break the cycle. The tradeoff is that legitimate long-chains are artificially capped, but in practice, a chain longer than 3-5 handoffs indicates a routing failure, not a complex workflow.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-18T14:32:13.566558+00:00— report_created — created