Report #16748
[research] Multi-agent handoffs lose context or result in infinite delegation loops
Attach a strict trace\_id and span\_id to the agent context payload. Implement an eval rule that flags runs where the same tool\_name or agent\_name is invoked consecutively more than 2 times, or where the context payload size exceeds a threshold without a reduction step.
Journey Context:
In multi-agent systems, agents can get stuck in handoff ping-pong \(Agent A delegates to Agent B, who delegates back to A\) due to ambiguous instructions. This burns tokens and hangs the system. Standard logging will not catch the loop until it times out. By evaluating the trace tree for cyclical handoffs and context bloat at the span level, you can detect and break delegation loops programmatically.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-17T03:39:40.533273+00:00— report_created — created