AI engineering
Making Multi-Agent Handoffs Clear and Recoverable
Multiple agents need explicit ownership and shared evidence. Design handoffs with bounded tasks, structured results, failure states, and a clear final reviewer.
Adding agents can divide work, but it also creates coordination costs. Anthropic describes orchestrator-worker patterns in which a coordinator delegates and combines results. In a business application, the hard part is often the handoff: what a worker is expected to deliver, which evidence it may use, and who decides whether the combined result is complete.
Give each worker a bounded responsibility
Define the input, deliverable, allowed tools, and completion criteria for each role. Avoid assigning overlapping write authority unless the system has an explicit conflict strategy. A worker asked to inspect records should return findings and source identifiers, while the component responsible for changes should own the actual mutation.
Use a shared task identifier and versioned artifacts. Passing a prose summary alone can lose qualifiers, uncertainty, and source references. A structured handoff can include observed facts, unresolved questions, proposed changes, and links to evidence. Keep instructions separate from untrusted material copied out of documents or tool responses.
Rehearse a realistic division of work
Imagine a hypothetical release assistant with workers for API compatibility, documentation coverage, and user-interface checks. Each worker evaluates a defined portion of the release and returns a result with evidence. The coordinator then checks whether their findings refer to the same build and whether any required area remains unreviewed.
Parallel work helps only where dependencies allow it. Documentation cannot reliably describe an API contract that is still changing without a synchronization point. Establish when workers may proceed independently and when they must wait for a stable artifact. If two findings conflict, preserve both until the coordinator resolves the disagreement using evidence.
Treat partial completion as a normal state
Workers may time out, lose access, or return an incomplete result. Define whether the coordinator retries, narrows the task, continues with a clearly labeled gap, or requests human input. Put limits on recursive delegation and repeated review cycles so the system cannot expand indefinitely.
Test a stale worker result, duplicate completion event, and coordinator restart. Store enough durable state to know which artifacts were accepted and which still need review. Measure overall task quality and elapsed time, not the number of agents or messages. More activity is useful only when it improves the final result the user receives.
