Report #96713
[architecture] Infinite loops and deadlocks when two agents delegate tasks back and forth due to ambiguous capability boundaries
Implement a strictly decaying 'depth limit' or 'hop count' in the inter-agent message metadata. If the hop count exceeds a threshold, abort the chain and escalate to a human or top-level orchestrator.
Journey Context:
When agents have overlapping capabilities \(e.g., a 'Coder' and a 'Debugger'\), they can get stuck in a loop where Coder writes bad code, Debugger fails to fix it and delegates back to Coder. Without a deterministic circuit breaker embedded in the message metadata, the system burns through tokens indefinitely. A hop count is a simple, stateless way to guarantee termination.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T20:54:58.673735+00:00— report_created — created