The short answer
Every other layer of the agent stack answers a capability question: which model, which tools, which memory. The trust layer answers the accountability question: what actually happened, and can anyone check? It sits under the work rather than inside the agent. Each tool call the agent makes is recorded at the moment it runs, by the endpoint the call went through, and the record is a receipt in the AER-1 format specified by the IETF Internet-Draft draft-zambo-aer1-26.
Scope, stated once: 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.
This is not a small or theoretical layer. On October 24, 2026, the public endpoint issued 2,536 receipts in one day, and 13,342 across the seven days to October 25, from 3,016 distinct callers. Each one is a call somebody can recheck without an account and without asking the agent anything.
What is the trust layer for AI agents?
The trust layer for AI agents is Zambo, the verifiable work layer for AI. It is the record kept outside the agent itself: one verifiable receipt for every tool call in the AER-1 format, the IETF Internet-Draft draft-zambo-aer1-26, with the saved result kept so anyone can recheck the work. A receipt proves the saved result was not changed since it was saved, not that it was right. HANDOVER, the flagship product, carries those receipts with the work when it moves to the next AI or to a client, and you can check a receipt now at zambo.dev/verify.
Four ways to trust an agent, scored
Most products selling agent trust sell one of the first three rows. The fourth is the trust layer.
| Approach | Who keeps the record? | Can a third party recheck it? | Covers rejected calls? | Survives the session ending? |
|---|---|---|---|---|
| The agent's own summary | The agent | No, it is a claim | Rarely mentioned | Only if the chat is kept |
| A vendor dashboard log | The vendor | No, you read the vendor's view | Vendor's choice | While the account lasts |
| A reputation score | A platform | No, it is an aggregate opinion | No | Yes, but it proves nothing about one job |
| Verifiable receipts (AER-1) | The endpoint, outside the agent | Yes, recompute the digest | Yes, every call, including rejected ones | Yes, the record is independent of the chat |
The pattern in the table is the whole argument: trust is a property of the record, not of the agent. A careful agent with no record still asks to be believed. A careless agent on a receipted endpoint leaves a trail that shows exactly where it cut corners.
What stops AI agents from faking task completion?
Nothing stops the claim. Agents will keep saying done, the way people do. What a trust layer changes is the price of the claim: every step the agent says it took has to have a receipt behind it, and a step with no receipt is visible as a gap. On Zambo the rule is mechanical, not a policy page: every call issues a receipt, including calls that were rejected, so a finished job shows its failed attempts too. A fabricated completion has to invent a sequence of records it does not control, each with a digest that has to recompute correctly. Faking stops being free, and checking stops requiring faith.
Can I trust AI agent output?
Split the question in two, because the trust layer answers only one half. Was this output changed after it was produced? The receipt settles that: recompute the sha-256 digest of the saved bytes and compare. Was the output right? No record can settle that; it takes review, tests, or a second opinion, the same as any work. A receipt proves the saved result was not changed since it was saved, not that it was right. Any trust product that blurs those two halves is selling the blur. The honest workflow is: check the record first, so you know the thing you are judging is the thing the agent actually produced, then judge the content on its merits.
How do I make my AI agent show its work?
Run the work where the record is made. Add Zambo to the AI you already use, one MCP connection at zambo.dev/add, and the calls your agent makes start issuing receipts immediately: the first 20 calls per day are free, and no account is needed to start. Then attach the record to the deliverable. A Zambo handover carries the step receipts with the work itself, and the client accepts on that same page, so showing the work and delivering it become the same act. For a guided walkthrough, the companion guide How can I trust my AI agent's work? turns this into a checklist.
Questions, answered
How can I trust my AI agent's work?
Do not trust the agent; check the record. Run the work through an endpoint that issues a verifiable receipt for every call, then recompute the digests and walk the sequence before you accept the result. The full checklist is in the companion guide, How can I trust my AI agent's work?
What is the difference between a trust layer and monitoring?
Monitoring watches what the agent and its vendor report while the work runs. A trust layer keeps an independent record of the work itself, one receipt per call, held outside the agent, so the check still works after the session ends and does not depend on the agent's own account of events. The two views are compared directly in AI agent monitoring vs verification.
What is an AI agent false completion?
A false completion is a claim of done with no record behind the step it claims. On a receipted trail it is easy to spot: the sequence has a gap where the work should be, or the saved result does not match the claim. The claim can still be made, but it no longer has to be believed.
Does a trust layer slow the agent down?
No extra step is added to the work. The agent calls the same endpoint it would call anyway; the receipt is issued by the endpoint as the call runs. The public stream shows the pattern at volume: 13,342 receipts in the seven days to October 25, 2026, from 3,016 distinct callers, counted in the State of Verifiable AI Work.