How do I get one receipt when one AI researches and another writes?
Keep both stages in one job on the same connection. The researcher's steps (sources consulted, states recorded) and the writer's steps (drafts produced from them) land in a single record in order, so the deliverable's receipt covers the whole chain: what was found, then what was written from it, checkable end to end.
Split pipelines break evidence at the handoff. The research happens in one tool, its output is pasted into another, and the writing tool's record, if any, begins from a block of unattributed text. When a client asks what a claim in the draft rests on, the chain is broken exactly at the joint where the answer lives.
The one-job pattern removes the joint. Both assistants work through the same MCP connection, https://zambo.dev/mcp , so their steps interleave in a single job in true order: research steps with sources and states, then the writing steps that consumed them, each labelled with its producer. The seam becomes a visible step of its own, the moment the research summary passed to the writer, which is the exact artefact a checker wants when a draft's claim looks unsupported.
The pattern generalises to any chain: planner and executor, drafter and reviser, data collector and analyst. The rule is constant: one deliverable, one job, every stage's steps inside it, producers labelled. Chains assembled that way produce one receipted object a client can accept whole, instead of fragments each tool vouches for separately and none vouches for together.
Checklist: the two-AI one-record job
1. One job opened for the deliverable, before stage one starts. 2. Research steps inside it: sources consulted, states recorded, summary step saved. 3. Handoff step: the exact research summary passed to the writer, saved as a step. 4. Writing steps inside the same job: drafts as versions, producer labelled per step. 5. Delivery: the job ships as one handover; the client accepts the whole chain in one decision. Test: pick any sentence in the final draft and walk it back to a source step without leaving the record.
Worked example
The shape is the one under any multi-step record, for example https://zambo.dev/handover/508a46e2-8d6e-4099-bd12-7ae461c72c55 , four steps in one job under one recipient. Extend the steps across two assistants instead of one and nothing structural changes, which is the point: the record does not care how many producers a job has, only that their steps land in it in order.
Frequently asked questions
The two AIs are on different platforms that cannot share a session. What then?
They do not need to share a session; they need to share the job. Each connects through the same MCP endpoint and appends its steps to the same job in sequence. The connection is the meeting point, not the chat window.
What if the writer ignores the research?
The record shows it: writing steps that cite none of the research steps are visible as a break in the chain, which the checker can see and the checker no longer has to take your word about. Visibility of the failure is the feature.
Does each AI get its own receipt instead?
Each step carries its own receipt inside the job, labelled by producer, and the job as a whole is the deliverable's record. You get both granularities: per-step proof, and one chain a client can accept.
By Brennan Zambo, founder of Zambo.
Last verified: 2026-10-17.
Try a handover
A handover gives your client one proof link with verifiable receipts to review, then Accept or Flag.
try it: say 'use zambo to hand this over to <client>' in your ai