In a multi-agent workflow, one of five parallel agents stops producing output and its assigned tasks remain incomplete. The other four agents have completed their tasks and are waiting. How should the system respond?

Degraded coordination in multi-agent workflows requires automated detection and graceful recovery mechanisms to maintain system throughput and reliability.

The correct response is the timeout or heartbeat check option: detect the stalled agent, terminate it, redistribute its tasks to available agents, and log the failure for investigation. A well-designed orchestrator implements heartbeat or timeout-based liveness checks for each agent. When an agent exceeds its expected completion time or stops sending heartbeat signals, it is declared stalled. The orchestrator terminates the stalled agent, or its execution context, logs the failure with diagnostic information, and redistributes the incomplete tasks to available agents. This approach recovers quickly without human intervention and maintains a complete audit record.

The indefinite-wait option is wrong because waiting indefinitely for a stalled agent blocks the entire workflow and may cause timeouts in downstream systems. In production, unresponsive processes must be actively managed, not awaited passively.

The cancel-and-restart option is wrong because canceling the entire workflow and restarting all five agents wastes the completed work of the four healthy agents. Recovery should be scoped to the failing component, not the entire workflow.

The mark-complete-and-submit-partial-results option is wrong because submitting partial results as complete is a dangerous deception. Downstream processes may operate on the assumption that all tasks were completed, leading to incorrect conclusions, missing data, or missed defects.