AI handoff
The transition pattern where an AI agent transfers control of a task to a human or to another agent — preserving context, state, and prior actions so the receiver can continue seamlessly.
Handoff is the highest-leverage pattern for production agents. An agent handles the work it does well, then hands off to a human (or specialist agent) when it hits a case it cannot. Done well, handoff captures the value of automation while keeping humans in control of the cases that need them.
The mechanics: the agent detects a handoff trigger (low confidence, sensitive case, explicit rule), packages the conversation context and state, and routes to the right handler. The receiver gets a structured summary, not a wall of logs. Good handoff design is what separates production-grade support agents from frustrating ones.
For agent operators, the handoff failure modes to design against: silent handoffs (the user does not know they were transferred), context-less handoffs (the human has to start over), and ping-pong handoffs (the human punts back to the agent which punts back). Each is a UX failure that erodes trust faster than any single wrong answer.
Where this shows up
Frequently asked
When should an AI agent hand off to a human?+
Low-confidence cases, sensitive topics (legal, medical, financial), explicit user requests, and any irreversible action over a defined threshold. Tune thresholds based on actual escalation patterns rather than guessing.
How does the human handler see the AI conversation history?+
Production-grade handoff packages a structured summary (intent, key facts, prior actions, current state) along with the full transcript. The summary is what humans read; the transcript is the audit trail.