AI Agent Handover: Pass a Live Job to Another AI Without Losing the Record

Somewhere around step seven of a ten step job, your AI hits a limit, or you realize a different model would finish it better, or the client asks the question that ends every chat workflow: "show me exactly what it did."

So you copy the chat into the next AI. What arrives is a summary. The summary is confident, tidy, and missing the one thing that mattered: the record. The figures the first AI actually pulled, the page it actually read, the step it quietly skipped. The second AI cannot tell what was checked and what was invented, so it either redoes everything or, worse, trusts the summary and builds on sand.

That gap is what an AI agent handover closes. A handover is not a summary you paste. It is the job itself, packaged as one link: every step in order, what each step produced, what was not checked, and a verifiable receipt for each saved result so anyone can confirm the record has not changed since it was saved.

Here is how to do one properly, start to finish.

Step 1: Do the work where steps get saved, not just said

A handover can only carry what was recorded. If your AI did the job inside a plain chat, there is nothing to hand over except its own narration.

Run the work through a tool layer that saves each call. In this walkthrough the job runs through Zambo's MCP endpoint (https://zambo.dev/mcp, added as a connector in whatever AI you use). Every tool call returns its result plus a receipt, a saved record you can open later at a public link. The receipt proves the saved result has not changed since it was saved. It does not prove the answer was right, and any tool that claims more than that is overselling.

Our example job is small on purpose: fetch the live BTC and ETH prices in USD, so a second AI (or a client) can continue from checked figures instead of a number typed from memory.

Step 2: Hand over the job, not a recap of it

When the work is done, or done enough to pass on, you create the handover. One instruction does it: "use zambo to hand this over to my client." What comes back is a single link.

Here is the real one from this walkthrough, created while writing this article:

https://zambo.dev/handover/d8de58c5-4f95-4ed0-8df0-186ad09ceea6

Open it. You will see the goal, the two steps that ran, the exact results they returned (BTC at $81,695.00 USD, fetched live at 09:01 UTC on October 23, 2026, and ETH at $2,404.07 USD, fetched at 09:02 UTC), and the receipt behind each step:

https://zambo.dev/run/9406b95f-0ec9-4a54-98a3-a020c4433391 (the BTC step)
https://zambo.dev/run/2d725e37-0154-4cf6-925f-4277ef18e96e (the ETH step)

One detail before you open them: the handover steps are redacted publication copies of these source receipts, and each one shows the result exactly as Zambo's live price tool recorded it, provider credit included. Nothing in either figure above was retyped or rounded by hand.

Anyone with the link can open the handover in a browser and check each step. No account, no wallet, nothing to install. The page is also honest about its own limits: it lists what a receipt does not prove (the outside timestamp, who issued the step, whether the result was correct). That honesty is the point. A handover that hides its gaps is just a summary with better formatting.

Step 3: The next AI picks it up and chains its work onto the same record

This is the part a pasted summary can never do. The receiving AI picks the job up from the handover (the page hands you the exact prompt to paste into it), reads the steps and their receipts as its starting state, and when it hands the job on again it chains off the original, so its handover links back to the first one instead of starting a disconnected thread. The trail reads as one chain: AI to AI to client.

In practice that means:

• The second AI does not re-run step one to "make sure." The receipt settles it: it proves the saved result for that step has not changed since it was saved, so the second AI starts from the record instead of from trust.
• If the second AI adds three steps, its handover carries them in order, chained to the record it picked up.
• Anyone reviewing later sees the whole chain, including the AI label reported for each part. Those labels are reported by the AI, not verified.

You can see a bigger delivery in this public example, five steps handed over for client review: https://zambo.dev/handover/bd7e5325-33fd-4c5e-b029-76b781fa2f51

Step 4: The receiver accepts or flags, on the record

A handover ends with a decision, not a shrug. The sender gives the client (or the next person responsible) their own client link: the public page anyone can open is read-only, and the client link is where they check the steps they care about and accept the delivery or flag the step that looks wrong. That decision is recorded on the handover itself.

Be precise about what that means: acceptance records that the client reviewed it and said yes. It does not prove the work was correct, and the handover page says so. What it kills is the follow-up argument. There is no "you never sent that version" when the version, its steps, and its receipts sit at one link that has not changed since it was saved. If a dispute does escalate, the handover can be downloaded as a dispute pack (PDF) with the full trail.

When to use a handover (and when not to)

Use one whenever a job outlives a single chat:

• You hit a model limit mid-job and the work has to move.
• A stronger or cheaper model should do the second half.
• A client, a teammate, or future you will ask "what exactly happened here?"
• Money, compliance, or a deliverable rides on the steps being checkable later.

Skip it for throwaway one-shot questions. If nobody will ever need to check the steps, a receipt is overhead.

Try it

Add https://zambo.dev/mcp to your AI as a custom connector, run any small job, then say "use zambo to hand this over." No connector setup? Run the same job in your browser at https://zambo.dev/chat, no account, then say the same sentence. The test handover on the handover page (https://zambo.dev/handover/) takes about ten seconds and needs no account. If the job matters, a Day Pass ($0.99 for 24 hours; agents pay in USDC on Base, humans check out by card at https://zambo.dev/day-pass/) unlocks handovers for client sign-off and the rest of the paid tools, so the job you just proved can keep running all day.

an agent receipt is a record of one ai agent tool call that anyone can check: it keeps the exact saved bytes and their sha-256 digest, and it proves the result was not changed since it was saved, not that it was right.

The habit to build is simple: never let a job travel as a summary again. Hand over the record, and let the next AI, or the client, check it themselves. Disclosure: I am rambo, director of ops at Zambo, and this walkthrough uses Zambo's own handover, linked above so you can check it yourself.