Agent Handoff
When a support call reaches someone who cannot help, the useful move is not to guess — it is to pass the caller to whoever can, along with everything said so far. An agent handoff is that transfer, performed between agents in the middle of a task.
How a handoff is wired
The common implementation makes the transfer itself a tool: the current agent is given a transfer_to_billing-style function, and calling it swaps in another agent's instructions and tool set for the remaining turns. The OpenAI Agents SDK models handoffs exactly this way, with a triage agent inspecting the request and routing it onward. What travels with the control matters as much as the routing — the conversation so far, the facts already established, and a stated reason for the transfer.
Handoff versus delegation
The two are easy to confuse and behave differently:
- A handoff gives up control — the receiving agent owns the task and answers the user directly.
- Delegation keeps it — a supervisor calls a worker agent as a subroutine and continues with the result.
Choosing the wrong one produces the familiar failures: two agents bouncing a task back and forth, or a specialist confidently answering the wrong question because the reason for the transfer never arrived with it.
Why it matters
Handoffs let a system be assembled from narrow, testable agents with small tool sets, instead of one agent carrying every instruction and every tool at once — which keeps prompts shorter, tool selection more accurate, and each part replaceable on its own.
An agent handoff transfers control of a task from one agent to another, passing along the context needed to continue. The receiving agent takes over the conversation or workflow with its own instructions and tools, and the first agent stops acting. Handoff is the lightweight alternative to a central orchestrator: specialisation is achieved by routing the task to the right agent rather than by one agent doing everything.
Ring a bank and the first person you speak to often cannot solve your problem. They pass you to someone who can — and, if they are good at their job, they explain your situation before they do. An agent handoff is that same move, between AI helpers.
Rather than building one program that knows everything, people build several, each good at one thing: one checks orders, another handles refunds. When a task turns out to belong to a different one, the first hands it over, along with everything it has learned so far.
Done well, you never notice it happened. Done badly, you get the software version of being transferred in circles.