Agent Handoffs That Don’t Lose Context

Single-agent prompts are easy to reason about. The pain starts when you split work across specialists: billing vs. technical support vs. compliance. The user shouldn’t feel like they’re starting over every time the “owner” of the conversation changes.

What usually goes wrong

• The second agent asks for information the caller already gave
• Summaries are vague (“they had a billing issue”) instead of structured
• Latency spikes because every handoff re-runs a giant prompt

A better contract

We treat a handoff as a typed event, not a free-text note. The outgoing agent attaches:

• Stated goal and current sub-task
• Facts the user has already confirmed (phone digits, account tier, language preference)
• Open risks (“user is frustrated”, “payment failed twice”)

The receiving agent gets that bundle as system context before the first reply, so the first sentence can acknowledge continuity: “Thanks for holding — I see the charge didn’t go through. Let’s fix that now.”

Design takeaway

If you’re building voice workflows, invest in the schema of what crosses the boundary. The transcript alone is rarely enough; you want a small, explicit state object that both agents agree on.