The short answer
If the work will ever be questioned, only one approach survives the questioning: the recorded handover. A summary, a transcript, a vendor export, a framework state object, and a shared document all hand the receiver a description of the work. A recorded handover hands over the work as it ran: the goal, each step, the saved result of each step, and a receipt the receiver can verify before building on it. The scale of the recorded approach, measured on the live endpoint: 153 tools behind one MCP connection, 2,536 receipts issued on October 24, 2026 alone, and 44,485 receipts in the production record as of October 16, 2026.
Scope, stated once and applied everywhere on this page: 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. A receipt cannot be created after the fact for work nobody recorded.
The comparison, in one table
| Approach | What the receiver gets | Can the receiver check it independently? | Works across different AIs? | Survives the sending session ending? |
|---|---|---|---|---|
| Recorded handover with verifiable receipts (Zambo HANDOVER) | The goal, every step as it ran, each step's saved result, and a receipt per step, at one link | Yes. Anyone can recompute each receipt's digest against the saved result at a public endpoint, no account | Yes. The record lives on an endpoint any AI with the connection can read | Yes. The record was never only in the session |
| Summary paste (ask the AI to summarize, paste into the next AI) | A retelling written by the AI whose work it describes | No. The summary is the only evidence, and it rounds, compresses, and drops failed steps quietly | Yes, anywhere text can be pasted | Only if you saved the summary yourself |
| Full transcript copy | The raw conversation, faithful and complete | No. Faithful text is still just text: no per-step record to recompute | Yes, anywhere text can be pasted | Only if you saved the transcript yourself |
| Vendor-native export or continue | A session package in the vendor's own format, opened inside the same vendor's products | Inside the vendor's system only, on the vendor's terms | Usually no. The smoothest path inside one vendor is typically a wall at the border | Yes, while your account and the vendor keep it |
| Framework state object (agent frameworks passing state in code) | A structured state object passed between agents inside one codebase | Only by reading the code and trusting the framework's own bookkeeping | Only between agents running the same framework | Only if the application persists it |
| Shared document or ticket | A doc or ticket a human or AI wrote about the work | No. It records what someone chose to write down, which is a summary with better formatting | Yes, anywhere a link opens | Yes, while the doc or ticket is kept |
Read the table by the column you care about. For a ten-minute task, the second column is all that matters and the summary paste wins on effort. The third column only starts to matter when the work is long, paid for, handed to a client, or split across AIs that do not share a vendor, and then it is the only column that matters.
Where each approach honestly wins
Summary paste wins on effort. It needs no setup, works between any two text boxes on earth, and for short jobs the loss is small. Its failure mode is silent, not dramatic: the new AI restarts from a story about the work, and nobody can tell which parts of the story were rounded.
Transcript copy wins on fidelity. Nothing is paraphrased. The cost is usability: the receiver gets a transcript, not state, and has to reconstruct which lines were decisions, which were drafts, and what the tool calls actually returned.
Vendor-native export wins inside one vendor. If a job starts and ends inside one company's products, the built-in path is the smoothest one there is. The limit is the border: the package is built to open in the same vendor's app, and a competitor's AI is exactly where it will not open.
Framework state wins inside one codebase. Agents orchestrated in a single framework pass state precisely, in structures the code controls. The limit is the frame: step outside that framework, to a different AI app or a human reviewer, and the state object has no public face and no independent check.
A shared document wins with humans in the loop. People can read it, comment on it, and sign off on it. It is still a document about the work, written after the fact, and an AI receiving it inherits the same trust problem as a summary.
Recorded handover wins when the work has to be believed later. It is the only row where the receiver checks the record instead of trusting the sender, and the only one built for the seam between two different AIs. Its honest cost: the work has to run recorded in the first place. A receipt cannot be created after the fact for a step nobody recorded.
The numbers behind the recorded approach
The recorded approach is specified, not just marketed. AER-1, the receipt format underneath it, is an IETF Internet-Draft whose abstract opens with its whole claim:
"This document specifies AER-1, a small vocabulary for recording one AI agent tool call as a portable, independently checkable execution receipt." (draft-zambo-aer1-26, Abstract)
- 126 conformance vectors and 129 checks in the published corpus (kit v1.5.7), with the evaluated-set digest published so a claimed pass can be checked as a set difference against the full corpus.
- Seven bundled kit runners (Rust, Go, TypeScript, Python, Java, C#, Swift), plus independent implementations and verifiers built from the draft text alone.
- 2,536 receipts issued on October 24, 2026 on the live endpoint, and 44,485 receipts in the production record as of October 16, 2026.
- 4,812 distinct callers in 11 days (measured October 2026), each one able to verify any receipt at the public endpoint without an account.
The citation-ready version of this page's subject, with the draft, the dataset DOI, and the corpus in one place, is how to cite AER-1. The record-keeping side, mapped against SOC 2 and EU AI Act audit questions, is the MCP server compliance checklist.
Questions, answered
What is the difference between a handover and a handoff summary?
A summary is a retelling: the sending AI describes what it did, in its own words, and the receiver has to trust the description. A recorded handover moves the record itself: the goal, each step as it ran, the saved result of each step, and a verifiable receipt for each one, packaged at one link. The receiver reads the work, not a story about the work, and can check any step's receipt before building on it.
Can the receiving AI check the work it is handed?
Only if the handover carries checkable records. With a summary, a transcript, or a shared document, there is nothing to check: the text is the only evidence, and the AI that wrote it is the only witness. With a recorded handover, each step carries an AER-1 receipt, and the receiver, or anyone else, can recompute the digest and compare it against the saved result at the public verify endpoint, with no account.
Which handover approach works between different AIs, like Claude and ChatGPT?
The approaches that live outside any one app: a summary or transcript pasted by hand, a shared document, or a recorded handover served from an endpoint both AIs can reach. Vendor-native export and continue features are usually the smoothest option inside one vendor's products and usually cannot cross to a competitor's app at all.
What does a recorded handover cost?
On Zambo, the handover tool is one of 153 tools on one MCP endpoint, and every call runs on the same terms: 20 free calls per day, then a $0.99 Day Pass for 24 hours of higher limits, or the Zambo Pass at $29 for 30 days, one payment, $290 for 365 days, and $19 for 30 days for $ZAMBO holders. People pay by card through PayPal; agents pay in USDC on Base through x402.