Updated
Design AI agent handoffs with useful case summaries, explicit ownership, durable state, approval expiry and recovery tests for human or agent takeover.
A reliable AI agent handoff transfers the task, relevant evidence and authority to a named receiver, then prevents the original agent from continuing conflicting work. Sending a transcript to a queue is only one part of that process. The system must know who owns the next action and what happens if nobody accepts it.
Imagine an order-support assistant that finds two refund requests for the same purchase. It can collect the order records, but a supervisor needs to decide whether the second request is valid. The handoff should explain the duplicate, link to the source records and stop automated refund execution. It should not make the customer repeat their complaint or describe an unapproved refund as completed.
Choose observable escalation triggers
Start with conditions that can be checked directly. “The model seems uncertain” is too vague to be the only rule, and a model's self-reported confidence is not a calibrated probability of correctness.
For this order workflow, useful triggers include:
The customer explicitly requests a person.
Two records conflict about the amount already refunded.
The proposed action exceeds the account's configured authority.
An identity check cannot be completed.
A required source system is unavailable beyond the recovery window.
The same tool failure repeats without new information.
The customer disputes a decision already made by automation.
A policy requires an authorized employee's approval.
Each trigger needs a destination and an expected response, not just a label. Identity failures may go to account support; duplicate payments may go to finance. Routing every case to one generic queue conceals operational problems until that queue becomes unmanageable.
Build a handoff packet the receiver can act on
Place the unresolved decision first. In the example: “Determine whether order 412 has already received its full refund; do not issue another refund until reconciled.” Follow with the verified facts, disputed facts and actions already attempted.
Separate the user's request from the agent's interpretation. Preserve the original message where wording matters, but do not require a supervisor to read an entire conversation before finding the problem. A short summary should link to source records and the relevant transcript section.
Include the current record version, time of last read and identifiers for completed side effects. If the agent sent an email, show the delivery record. If it only drafted an email, state that explicitly. Missing this distinction can cause a human to send the same message twice.
The receiver also needs the permitted next actions. “Review case” is less useful than “approve no refund, approve the remaining amount, or request payment reconciliation.” Present evidence neutrally rather than steering the reviewer toward the model's preferred answer.
Capture the final decision and correction reason in a structured field. Review recurring correction patterns for additions to the evaluation set; do not automatically train on every reviewer click. Audit a sample for rushed approvals and avoid queues so overloaded that reviewers cannot inspect the evidence.
Transfer ownership atomically
Use a durable case record with an owner, state and version. When a human accepts the case, update ownership before enabling their action controls. The agent should check that same record before any later write.
Consider the race where a supervisor accepts the case while the agent's retry is still running. A final authorization and ownership check at the action boundary can prevent a second refund even if both processes started legitimately. Disabling a button in the chat interface is insufficient because a background worker may still hold an old job.
Keep customer-visible state aligned with the case record. A queued case can say it is waiting for review; an accepted case can identify the responsible team. Do not invent a waiting-time promise that the operational queue cannot support.
Persist enough state to resume safely
Store the information needed to recover after a worker restart: case identifier, workflow step, source versions, prior action receipts and pending decision. Avoid depending on a process's in-memory conversation history.
For teams using LangGraph, its interrupt documentation describes pausing execution for external input. Its persistence documentation explains checkpointed state and threads. These facilities support a handoff; they do not define your authorization policy or guarantee that replayed side effects are safe.
Check the framework's resume behavior before putting actions around an interruption point. An operation executed again during recovery must either be safe to repeat or protected by a stable action identifier. Keep the payment receipt outside an ephemeral model message so a restarted workflow can reconcile it.
Bind approvals to a specific proposal
A supervisor approving a refund of one amount has not approved a later, larger amount. Bind approval to the action parameters, relevant record version and intended account. Expire or invalidate it when those inputs change.
Re-read material facts immediately before execution. If another employee issued the refund after the handoff packet was created, the old approval should not authorize another payment. Show the changed state and return the case for a fresh decision when necessary.
The tool reliability guide explains why schema validation cannot replace business validation. An action can be perfectly formatted and still be inappropriate for the current order.
Test the queue, not only the conversation
Run end-to-end scenarios with the actual staff interface and background workers. Measure how much reconstruction the receiver needs and whether the case reaches a verified outcome.
Ask for a person before the agent has gathered any details; check that the request is honored.
Restart the worker while the case waits for review; confirm the case remains actionable.
Have two reviewers accept simultaneously; verify that only one obtains ownership.
Change the order after approval; confirm stale approval cannot execute.
Leave the queue unstaffed; check that escalation and user messaging follow the agreed policy.
Return the case to automation; ensure the human's decision and limitations survive.
Use agent acceptance testing for action-level evidence and incident response planning when a failed handoff has already affected customers.
Frequently asked questions
Must every handoff include the complete transcript?
No. Provide a concise decision packet plus permission-controlled access to relevant history. Sending everything can bury the important facts and expose information the receiver does not need.
Can another agent be the receiver?
Yes, but ownership, permissions and completion criteria still need to be explicit. Record whether the second agent may act independently or only return a recommendation for the original workflow.
How should handoff quality be measured?
Track time to acceptance, time to resolution, repeated customer questions, reopened cases and duplicate actions. Review a sample of packets for factual accuracy. A low escalation rate alone can hide cases where the agent should have stopped but continued.
Written by
Moueen Togarvi
Founder & CEO at Voquarn Code, focused on product engineering, search growth, and practical AI systems.
