Skip to content

Handoffs

Stage 10 - Multi-Agent Systems

A handoff happens when one agent transfers control of a task or conversation to another agent. The first agent decides, "This is no longer my job," and the next agent takes over with the right instructions, context, and tools.

Transfer control Specialist takes over Sequential workflow Clear ownership Context matters

Goal

Understand handoffs as a beginner-friendly multi-agent pattern.

After this lesson, you should be able to explain:

  • what a handoff is,
  • how it differs from calling a tool,
  • when one agent should hand off to another,
  • what information should be passed during a handoff,
  • what can go wrong,
  • how to design a simple handoff flow.

Before You Start

Think about customer support.

You call support.
Front desk listens to your problem.
They realize it is a billing issue.
They transfer you to the billing specialist.
The billing specialist continues the conversation.

That transfer is a handoff.

In an agent system:

General support agent
  -> decides this is a billing problem
  -> hands off to billing agent
  -> billing agent owns the next turn

The important rule:

After a handoff, the receiving agent becomes responsible for the task.

Part 1: The Core Idea

A handoff is not just asking another agent a question.

It is a change in ownership.

Simple definition:

A handoff transfers control from one agent to another agent because the next
agent is better suited to continue the task.

Simple Picture

flowchart LR
    U[User request] --> A[Front desk agent]
    A --> D{Right specialist?}
    D -->|Billing| B[Billing agent takes over]
    D -->|Technical| T[Technical agent takes over]
    D -->|General| A

How to read this diagram: the first agent receives the request, decides which specialist should own it, and transfers control.

Part 2: Handoffs vs Supervisor-Worker

Handoffs and supervisor-worker systems can look similar, but the ownership is different.

Pattern Who owns the user task? What happens after another agent is involved?
Supervisor-worker Supervisor keeps ownership Worker returns result to supervisor
Handoff Receiving agent gets ownership New agent continues the task

In a supervisor-worker system:

Supervisor asks researcher for facts.
Researcher returns facts.
Supervisor continues.

In a handoff:

Support agent transfers to billing agent.
Billing agent continues with the user.

Beginner shortcut:

Supervisor-worker means "help me with this subtask."
Handoff means "you take over from here."

Part 3: What To Pass During A Handoff

A handoff should carry enough context for the next agent to continue without making the user repeat everything.

Useful handoff data:

  • the user's original request,
  • a short summary of the conversation so far,
  • why the handoff is happening,
  • the user's goal,
  • important constraints,
  • already completed actions,
  • open questions,
  • permissions or safety limits,
  • links to relevant records or tool outputs.

Example handoff packet:

{
  "from_agent": "front_desk",
  "to_agent": "billing",
  "reason": "User is asking about a refund status.",
  "user_goal": "Find out why refund has not arrived.",
  "summary": "User ordered item A102 and requested a refund 8 days ago.",
  "known_facts": {
    "order_id": "A102",
    "refund_requested": true
  },
  "open_questions": [
    "Has the refund been processed by the payment provider?"
  ]
}

Without this context, the receiving agent may ask duplicate questions or make wrong assumptions.

Part 4: When To Use Handoffs

Handoffs work best when a task moves through clear stages or domains.

Good use cases:

  • customer support routing,
  • onboarding flows,
  • sales to technical support,
  • triage to specialist,
  • legal review after draft generation,
  • human approval after automated preparation,
  • incident response stages.

Handoffs are especially useful when different agents need different tools.

Example:

Agent Tools
Front desk agent account lookup, ticket creation
Billing agent invoice lookup, refund status
Technical agent logs, diagnostics, deployment status

The front desk agent should not need every billing and technical tool. It only needs enough information to route the user.

Common Beginner Mistakes

Mistake What Goes Wrong Better Choice
Handoff too early User gets bounced around Let the first agent collect basic facts
No handoff summary New agent repeats questions Pass a short structured summary
Unclear ownership Both agents try to answer Define who owns the next turn
Too many handoffs Conversation feels broken Limit handoffs and use clear routes
Missing return path Specialist gets stuck Allow escalation or clarification

End Example: Support To Billing Handoff

User says:

I returned my headphones last week, but I still do not see the refund.
Can you check what happened?

Step-by-step handoff:

1. Front desk agent receives the message.

2. It identifies the topic:
   "This is about refund status, so billing should handle it."

3. It collects or confirms basic information:
   - order ID
   - return date
   - user's account

4. It creates a handoff packet:
   - reason: refund status
   - known facts: order ID and return date
   - open question: whether refund was processed

5. Billing agent takes over:
   "I can help with the refund. I see the return was received on Monday.
   Let me check the payment processor status."

The key idea is simple:

The first agent does not finish the task.
It transfers the task to the agent that should own it.