Handoff

OpenAI's Agent SDK takes a more engineering-oriented approach, natively supporting agent-to-agent hand-off mechanisms and built-in safety guardrails.

Agent Handoffs allow control to be transferred from one agent to another specialized agent during a conversation.

The Swarm framework has only two core concepts—Agent and Handoff—adopting a stateless orchestration design philosophy that completes inter-agent collaboration within a single function call chain.

The cost of Handoff is that triage cannot merge multiple answers; the cost of Agent-as-Tool is that the expert only knows what is in the task string, and if the string omits details, the expert has no way to know them.

Handoff mode suits scenarios where a single expert takes over, with the receiver owning the response, writing the reply, and handling subsequent interactions; Agent-as-Tool mode suits scenarios where multiple experts' opinions need to be synthesized into one coherent reply, with the caller retaining ownership and replying uniformly after aggregating the answers.

Each Handoff adds exactly 2 messages (the fixed pattern of 11-9=2 and 7-5=2 in the demonstration).

In the Priya refund case demonstrated in the article, a single agent handling the refund requires 9 messages; with Handoff, the refunds run contains 11 messages, the extra 2 being triage's transfer request and the "transferred to refunds" result; if triage's system prompt is retained, the total is 12 messages.

When a Handoff occurs, the receiving agent starts a new run with its own system prompt and inherits the prior user messages and handoff record, but does not inherit the original agent's system prompt (unless explicitly passed).

In Handoff mode, after the runtime recognizes stop reason handoff, it no longer returns control to the original agent but starts the receiving agent and inherits the conversation history, so the original agent's output is empty.

Apple's Handoff and Microsoft's Phone Link offer relatively stable cross-device collaboration experiences, thanks to end-to-end control over hardware and the protocol stack.

Apple often adopts a pattern in product feature design of first launching a technical capability and then waiting for user scenarios to grow naturally, a strategy that appeared in the early stages of features such as AirDrop and Handoff.

Handoff is a runtime primitive in multi-agent systems that reorganizes the conversation history, rebinds the tool set, and decides which state should be retained; poor design leads to privilege creep.

The Handoff pattern is suitable for one-time transfer scenarios, such as a customer service bot completely handing over a complaint case to a dedicated handling team

During a Handoff, the current message history is passed along to the new Agent, without requiring the caller to explicitly manage state transitions.

The Handoff mechanism transfers control by having an Agent function return another Agent object, underpinned by OpenAI's Function Calling capability.

Handoff allows the receiving Agent to have a completely independent system prompt and toolset, achieving separation of responsibilities.

In openai-agents-python, Handoff is designed as a first-class citizen, with each Agent able to declare which downstream Agents it can hand off to.

Handoff refers to one Agent, after completing its own responsibilities, passing control and context information to another Agent better suited to handle subsequent tasks.

The previous Handoff feature was primarily used to hand off tasks seamlessly between iPhone, iPad, and Mac.

Handoff uses Bluetooth Low Energy (BLE) for device discovery and iCloud for state synchronization.