Agent handoffs

When multiple agents participate in a task, handoff is the moment one passes control to another. Done cleanly, it's invisible to the user. Done poorly, context is lost and the user has to repeat themselves.

The handoff payload

What passes from agent A to agent B:

• Original user goal
• Context gathered so far (facts, decisions, partial work)
• Specific task for agent B
• Constraints (budget, time, quality bar)
• Success criteria for the handoff

When to hand off

• Current agent has reached its specialty boundary
• Another agent has better tools for the remaining work
• Escalation to a more capable model (cheap → expensive when needed)
• Human-in-the-loop (agent → human)

The clean handoff pattern

{
  "original_goal": "User wanted X",
  "progress_summary": "I've done A, B, C. Remaining: D, E.",
  "next_task": "Do D using tools X, Y",
  "context_to_preserve": [...key facts...],
  "constraints": {"budget": "$0.50", "timeout": "60s"}
}

The user experience

User shouldn't know they were handed off. The receiving agent introduces itself naturally: "I'll take it from here on the technical review." No "transferring you to a different department" moments.

Failure modes

• Handoff loses critical context → user has to repeat
• Both agents partially do the same work → wasted cost
• Handoff happens too late → receiving agent can't catch up
• Handoff happens too early → sending agent hadn't finished its part

Designing the trigger

Make handoff conditions explicit in each agent's system prompt. "When X happens, hand off to agent Y via tool Z." Clear boundaries reduce thrashing.