Skip to content
← All versions
v1current

Hand off between two agents without losing state

Define what the first agent must pass to the second, and prove nothing needed was dropped.

by AMARA · published Sep 13, 2026

Multi agent handoffs fail quietly. The second agent receives a summary, the summary omits a constraint the first agent had discovered, and the second proceeds confidently without it. This defines the handoff as an explicit contract rather than a conversational summary, and tests it by checking whether the second agent can answer questions that require the dropped information.

Target models
claude-opus-4.5, gpt-4o
Time
40 min
Difficulty
Hard
Cost
9,000 input tokens, ~2,800 output tokens, plus replay

Tools / stack

Claude API with tool useyour orchestration layera test harness

Inputs required

Both agents’ system prompts, an example handoff that went wrong, and the task the second agent must complete

Steps

  1. 1.List what the second agent must know to do its job, derived from its task rather than from what the first agent happened to produce.
  2. 2.For each item, state whether it is a fact, a constraint, or a decision already made. These carry differently and constraints are the ones usually lost.
  3. 3.Define the handoff as a structured object with a field per item, not a prose summary.
  4. 4.Make any missing required field a hard failure at handoff time, rather than something the second agent discovers later.
  5. 5.Test by asking the second agent questions that can only be answered from the handoff. Anything it cannot answer was dropped.
  6. 6.Replay the failed handoff through the contract and confirm the missing item is now caught at the boundary.

Expected output

structured handoff contract, a list of required fields with their type, and a replay showing the original failure caught at the handoff

Golden Example

Input
A research agent handing to a drafting agent, plus a run where the draft ignored a jurisdiction constraint the research agent had found
Output
Contract with nine required fields, three typed as constraints. The jurisdiction constraint became a required field. Replay failed at the handoff boundary with a named missing field rather than producing a wrong draft.