How a Team Hands AI Work Between People and Models Without Losing the Thread

Every team running agents has the same scar tissue. Ana starts a client job in one AI on Monday morning. The session dies, or the model gets swapped, or her shift ends, and Ben picks the job up on Tuesday from a paragraph Ana wrote about the work instead of from the work itself. He re-runs two steps that were already done, because nothing in the paragraph proves they ran. He skips one that was not done, because the paragraph implied it was. By Friday the client deliverable is fine and the record of how it got made is fiction.

The failure is not effort or tooling. It is that the job lived in a session, and sessions are the least durable object in the building. People change shifts. Models hit limits. Contexts get compacted. The job is the thing that has to survive all of that, so the job needs a home that is not a chat window.

The pattern: wrap, pick up, continue

On Zambo the job lives in its receipt chain, and moving it is a first-class operation called a handover. The wrap takes four things: the goal in its original words, a summary of where the work stands, the named recipient, and the next steps as an explicit list. A bare wrap is refused. That refusal is the design: a handover without a recipient and next steps is just a summary with ambitions.

Pickup is the other half, and it is always free, in any AI, because charging for the moment a job changes hands would tax exactly the behavior that keeps work continuous. The receiver does not get a retelling. It gets the goal, the state, the next steps, and the chain of receipts from the steps already run, each one recomputable. It continues the same job instead of starting a tribute band version of it.

One real wrap and pickup, on the record

This is not a diagram. Here is a handover from November 6, a two-asset price comparison one agent started and a second agent picked up:

https://zambo.dev/handover/5700bcf5-3f27-4c19-8b42-71a8c77e5677

Read what the page actually proves, because it is the right thing to prove. The wrap is there: the goal in its original words, the finished steps with their receipts, the next steps written down. Below it sits a continuation entry from a second agent, with a receipt of its own, timestamped that afternoon at 3:54 PM EST. The early steps are redacted publication copies that still show their saved results word for word. The supplied work behind the proof-loop steps is not republished. What the page hands everyone is the chain of custody: who held the job, what state it was in, who picked it up, and when, in an order nobody can quietly rearrange. For a team, that custody chain is the shared record. The contents live in the job. The custody is what everyone can check, and it is the part every substitute loses.

Why the receipt chain is the shared record

Teams try three substitutes first. A shared doc drifts within a day, because editing it is free and nothing marks which lines describe finished work. A ticket comment thread records what people said about the job, in between jokes and status emoji. A summary written by the outgoing agent has the deepest flaw: the agent whose work is being handed over is the one describing it, and the receiver cannot separate a faithful summary from a confident one.

Receipts have none of these problems because they are not written about the work. They are written by the work: caller, tool, saved result, timestamp, sequence position, and a SHA-256 commitment over the result bytes, fixed at the moment each call returned. On this page the first steps are redacted publication copies that still show their saved results word for word. What is not republished is the supplied work behind the proof-loop steps. The chain recomputes either way. Across 153 tools behind one endpoint, 36 on the default surface, every step of a job lands in the same shape, so a job that crosses three people and two models still reads as one job. The AER-1 format behind the receipts is an open draft with a public corpus of 143 vectors and 146 checks and five independent implementations, which is what makes the chain a record instead of a vendor's diary.

The economics of not losing the thread

The free tier carries 20 calls plus one wrap-up every day, pickup stays free, a Day Pass at $0.99 makes a day unlimited, the Pass runs $29 for 30 days ($19 for $ZAMBO holders), and Lifetime is $250 once. Against that, price one bad handoff: a senior re-running a morning of agent work, or a client deliverable rebuilt because nobody could prove which version of a step produced it. The wrap costs less than the coffee cooling next to it.

Keep the job in the record and the team can change everything else: the model, the shift, the person, the day. Your AI said it did the work. The next pair of hands should not have to take its word for it.