AI Agent Handoffs: How Humans and AI Should Transfer Tasks Between Each Other

An AI agent handoff transfers an in-progress task between a person and an agent, or between two agents. A good handoff moves more than the conversation: it carries the goal, current state, what was done and decided, evidence, open questions, who owns the task now and what the receiver is permitted to do.

Design three directions separately. Human → AI is delegation: scope and permissions matter most. AI → human is escalation: context and urgency matter most. AI → AI is orchestration: structured inputs, ownership and audit matter most.

Context transfer means the receiver understands the situation: who the customer or requester is, what they want and why. State transfer means the receiver knows exactly where the task stands: which steps are complete, which are pending, which records were created. Ownership must move explicitly; a task with two owners has none. Task status tells everyone whether it is in progress, waiting or blocked. Permissions may change at the handoff: a human supervisor may approve what the agent could not, or a specialist agent may need different access. Conversation history should be available, but summarized; nobody should have to read 60 messages. Evidence (documents, tool results, screenshots) shows why decisions were made. Incomplete work must be labelled as such, so a half-done refund is not mistaken for a finished one. Escalation details say why the handoff happened and how urgent it is.

Delegation fails when the agent's mandate is vague. The interface for handing work to an agent should capture the goal in the user's words, the scope (which accounts, which date range), the limits (spend, recipients, actions allowed), when to ask for approval and when the task should be considered done. Show the user a short plan before the agent starts on anything non-trivial, and make the task visible afterwards in a task list with status. For long-running delegated work, see background AI agent UX.

The most visible failure in AI support is the customer who explains their problem to a bot, gets transferred, and explains it again. Prevent it by generating a handoff record (as above) and showing it to the person receiving the task before they reply. Tell the customer honestly what happens next and when. Route by reason and urgency, not just to a general queue; AI customer support automation covers routing in support teams. In commerce protocols the same idea appears as a formal escalation state: the UCP checkout specification, for example, requires the merchant to return a continue_url when a checkout needs buyer input the agent cannot provide, so the buyer can finish on the merchant's site (UCP checkout).

When one agent hands work to another, use a structured task message rather than free text: goal, inputs, constraints, permissions, expected output format and deadline. The receiving agent should run with its own scoped permissions, not inherit everything the sender had. Keep a single owner (usually the orchestrator) responsible for the overall task, and carry one trace ID through every hop so the audit trail is complete. See agent-to-agent communication and AI agent orchestration.

Treat handoffs as first-class objects in your system, not as side effects of a chat.

Useful signals: how often customers repeat information after a handoff, time from escalation to first human action, share of handoffs returned because context was missing, tasks that stall with no owner, and resolution rate after handoff. Review a sample of handoff records each week; they show exactly where the agent's scope or context is wrong.

Handoffs are where human-AI systems succeed or frustrate people. Treat each one as a structured transfer of goal, state, evidence, ownership and permissions, design the three directions separately and make tasks visible in a shared store with a single owner. For when an agent should stop and ask instead of handing off, see AI action confirmation UX.