Agent Handoff
Synonyms: human handoff, live agent transfer, escalation, bot-to-human handoff, agent-to-agent delegation
Definition
Use cases
- The repeat-yourself transfer: A customer explains a double charge to a bot over 6 messages. The human who joins opens with "How can I help you today?"
- The escalation loop: The bot can't answer, asks "Did that help?", then shows the topic menu again. The user types "agent" 4 times and closes the tab.
- Vague briefs between AI agents: Anthropic's early research system sent subagents one-line tasks like "research the semiconductor shortage." The agents duplicated each other's searches and left gaps.
How it's used in practice
- List the triggers first: Hand off when the user asks for a person, after 2 failed answers, on payment issues, and on angry language.
- Send a summary with the transcript: The issue, the account, what the bot tried, and what the user wants. Copilot Studio passes the last topic and the user's phrases to the live agent automatically.
- Tell the user what's next: Name who takes over and how long the wait is.
- Brief AI agents in full: Each brief needs a goal, an output format, the tools to use, and scope limits.
Challenges & limitations
- Context gets cut: Long transcripts get trimmed between agents, and key details go missing.
- Empty queues: A handoff at 2 a.m. often goes to a queue no one is watching.
- Wrong success metric: Teams that track "containment rate" reward bots that avoid escalating, even when the user needed a person.
Free resources
- Handoffs (OpenAI Agents SDK) — official docs on delegating tasks between agents and filtering conversation history
- Building Effective Agents (Anthropic) — when to use agents at all, plus human checkpoints and stopping conditions
- Hand Off to a Live Agent (Microsoft Copilot Studio) — the exact context variables a bot passes to a human agent
