An agent handoff is the transfer of control from one AI agent to another during an agentic workflow. It is used when a specialist agent is better suited to continue the task than the current agent.
For example, a triage agent may hand off a billing question to a billing agent, a code issue to a coding agent, or a workspace-heavy task to an agent sandbox.
An agent handoff normally defines:
- the destination agent;
- a description of when the handoff should be used;
- optional structured metadata such as reason, priority, or language;
- an input filter that controls what history the receiving agent sees;
- approval or guardrail behavior; and
- tracing metadata for debugging.
Handoff is different from calling a sub-agent as a tool. In a handoff, the receiving agent becomes responsible for the next turn of the same top-level run. In an agent-as-tool pattern, the outer agent remains in control and the specialist executes as a nested operation.
The main risk is context leakage or context loss. Passing the full conversation may expose irrelevant or sensitive information. Passing too little may omit constraints needed for correctness. Context engineering and explicit input filters are therefore important.
Handoffs should be evaluated with realistic routing cases. A handoff description that sounds clear to a developer may still be ambiguous to the model. Incorrect routing can silently degrade quality, increase cost, or bypass needed controls.
The OpenAI Agents SDK handoffs documentation describes handoff descriptions, structured handoff input, input filters, and history management.
The LLM Knowledge Base is a collection of bite-sized explanations for commonly used terms and abbreviations related to Large Language Models and Generative AI.
It's an educational resource that helps you stay up-to-date with the latest developments in AI research and its applications.