Report #71742
[architecture] Infinite routing loops where agents bounce a task back and forth due to low confidence
Implement a monotonically decreasing routing budget or strict escalation path. When an agent routes a task due to low confidence, the router must either increase the capability tier of the next agent or decrement a global routing counter to force a terminal action.
Journey Context:
Confidence-aware routing \(where an agent passes a task to a 'better' agent\) seems elegant but creates a classic distributed systems livelock. If Agent A thinks Agent B is better, and Agent B thinks Agent A is better, the task loops indefinitely. LLM confidence scores are also poorly calibrated. By enforcing a strict escalation DAG \(directed acyclic graph\) or a hard limit on handoffs, the system guarantees termination, forcing the highest-tier agent or the last-resort agent to handle the task even if uncertain.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T03:00:22.539166+00:00— report_created — created