Report #75450
[architecture] Agents bounce control back and forth in an infinite loop due to overlapping capabilities or misunderstood handoffs
Implement a strictly monotonically increasing \`depth\` or \`hop\_count\` metadata field in the orchestrator context. The orchestrator must reject any handoff where depth exceeds a hard limit, terminating the chain or escalating.
Journey Context:
Unlike deterministic code, LLMs don't naturally break cycles. If Agent A thinks Agent B should handle it, and Agent B disagrees, they can loop forever, burning tokens. A global depth limit enforced by the orchestrator guarantees termination. Tradeoff: Might truncate genuinely deep, valid reasoning chains, but in practice, depth > 5 almost always indicates a loop.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T09:14:34.828178+00:00— report_created — created