Agent Handoff

Agent handoff is the pattern where one LLM agent delegates a conversation or task to another agent better suited to handle it — analogous to a customer-service representative transferring a call to a specialist — and is foundational to multi-agent systems, agent-supervisor architectures, and modern conversational AI applications. The handoff carries context: the new agent receives the conversation history, the user's underlying intent, any structured state accumulated so far, and explicit instructions about what the previous agent was unable to do. Handoff is sometimes called "agent routing," "agent transfer," "delegation," or "escalation" depending on the framework and direction (peer-to-peer vs upward to a more capable agent). OpenAI's Swarm framework (October 2024) and Agents SDK (March 2025) made handoff a first-class primitive — agents have a `handoffs` field listing other agents they can transfer to, the LLM chooses to handoff via a tool call, and the framework swaps the active agent while preserving conversation state. LangGraph supports handoffs as conditional edges between agent nodes. Anthropic's Claude supports handoff patterns via tool-use orchestration. The practical engineering challenges: deciding when to handoff (training the LLM to recognize "I can't help with this"), passing the right context (too little and the new agent restarts from scratch; too much and the context window bloats), and preventing handoff loops (Agent A hands to B, B hands back to A). Production handoff systems include explicit handoff policies (rule-based or learned), conversation-state schemas that survive handoffs, and audit logging at every handoff boundary. A common pattern in customer-service AI: the triage agent classifies the issue, hands off to a specialist agent (billing, technical support, legal, escalation to human), and the specialist may itself hand off based on what it discovers. AI governance teams treat handoffs as critical audit events because the responsibility for the conversation literally changes — a different agent's policies, capabilities, and risks now apply.

Handoff discipline from 25 years of workflow transfers: Centralpoint has tracked workflow transfers — content from author to reviewer to publisher, ticket handoffs between teams — for 25 years with full audit lineage. Agent handoffs inherit that same handoff discipline. Handoffs run on-premise, tokens meter per skill (and per handoff), and handoff-orchestrated chatbots deploy through one line of JavaScript.


Related Keywords:
Agent Handoff,Agent Handoff,Oxcyon, AI, AI Governance, Generative AI, Inference, Inference, Inferencing, RAG, Prompts, Skills Manager,