Aglet

Investigate a Failed Agent Handoff

Handoff failures often appear as vague escalation, duplicated work, or a user who does not know who owns the next step. Investigate the transfer as a sequence. Compare what the agent knew with what the recipient received, and preserve the exact task state at handoff.

Build a useful investigation brief

  1. Reconstruct the boundary

    Record request, agent state, tool results, uncertainty, trigger condition, transfer message, packet fields, recipient, and expected next action. Capture whether the user accepted or understood the transfer. Preserve versions and task fixture.

  2. Compare recipient readiness

    Have a reviewer follow the received context without hidden history. Mark missing decisions, evidence, constraints, authority boundaries, or links to completed work. Separate a deficient packet from a recipient process that was never represented in the test.

  3. Test alternate paths

    Replay with clearer context, earlier escalation, and a case that should remain with the agent. Change one trigger or packet field at a time. State whether the fault is detection, summarization, ownership, user notice, or unobserved recipient behavior.

What to carry forward

The investigation brief should link the trigger, state, packet, recipient reading, and replay. End with a localized handoff failure or evidence gap. Do not blame the transfer for recipient behavior that the evaluation never measured.

Technical background: Google DeepMind evaluation research.

Keep the decision with the work.

Use a Work Item in Aglet to record the problem, the evidence you have, and the next decision. Add an owner and priority, then keep updates in the discussion so the next person can follow the reasoning.

Create an account See the product workflow