Async agent callbacks reduce correct task completion in a controlled benchmark

An agent can receive a callback after its assumptions about state, permissions or prior effects have changed. A delayed message may be duplicated or arrive after a different actor has updated the same resource. Without explicit version checks, a correct-looking continuation can repeat an effect or apply a stale decision.

The authors evaluate 200 executable tasks under eight asynchronous disruption scenarios with nine agents, reporting average task success of 26.0 percent compared with 29.7 percent for the paired batched condition. The study uses controlled tasks and particular frameworks; the performance difference is not a universal comparison of async versus sequential design and not a quantified production incident rate.

Bind callbacks to task instance, state version and effect receipt. Reconcile an unknown prior outcome against external evidence before retrying or committing a new action. The lasting issue is not latency but the separation between an agents remembered plan and the actual current authority and effect state. Primary evidence.

Additional evidence and trust boundary

A callback should not be judged only by whether the final artifact is eventually correct. The contract must also protect state that changed while a subagent was running, prohibit repeated effects, and require re-verification after a conflicting event. Task-level success can therefore coexist with an unauthorized mutation. The paper reports a benchmark disadvantage for the tested asynchronous delivery policy, not a reason to avoid asynchronous execution categorically. Supporting record.