Agent Handoff
An agent handoff is a pattern in multi-agent systems where one agent transfers control of the conversation or task to another agent — passing along context but ceding ownership of the loop. It differs from delegation, where the original agent retains control and consumes the delegate's output.
Handoffs are used in OpenAI's Agents SDK, in the earlier Swarm prototype, and in handoff-style multi-agent designs more broadly. Handoffs reduce the orchestrator's reasoning load by routing whole sub-tasks to specialists, but they require precise instructions on each agent about when to hand off versus continue.
Example
A triage agent receives an inbound customer message, classifies it as a billing issue, and hands off to a billing-specialist agent — which then owns the rest of the conversation, including any further handoffs. The triage agent's instructions explicitly enumerate the handoff destinations and when to choose each, so the routing decision is grounded in the prompt rather than inferred from broader model behavior.
Frequently asked questions
What is Agent Handoff?
- An agent handoff is a pattern in multi-agent systems where one agent transfers control of the conversation or task to another agent — passing along context but ceding ownership of the loop. It differs from delegation, where the original agent retains control and consumes the delegate's output.
How does Agent Handoff work?
- Handoffs are used in OpenAI's Agents SDK, in the earlier Swarm prototype, and in handoff-style multi-agent designs more broadly. Handoffs reduce the orchestrator's reasoning load by routing whole sub-tasks to specialists, but they require precise instructions on each agent about when to hand off versus continue.
Can you give an example of Agent Handoff?
- A triage agent receives an inbound customer message, classifies it as a billing issue, and hands off to a billing-specialist agent — which then owns the rest of the conversation, including any further handoffs. The triage agent's instructions explicitly enumerate the handoff destinations and when to choose each, so the routing decision is grounded in the prompt rather than inferred from broader model behavior.