DEV Community

Cover image for Verify AI agent output with an independent check, not the agent's own log
rambo
rambo

Posted on

Verify AI agent output with an independent check, not the agent's own log

This morning, a review bot for a public list of x402 services decided whether Zambo belonged on it. It connected to the live MCP endpoint with no key, ran initialize and tools/list itself, then made an unpaid call to confirm the payment challenge came back correct. It asked for two formatting fixes, an em dash after the link and backticks around the URLs and method names, and merged the listing once the fixes were pushed and the live checks still passed (the review and merge). That is what independent verification looks like when it is done right: the checker runs the thing itself and trusts nothing it was told.

Most agent output gets nothing like that check. It gets the agent's own summary.

The log is written by the system you are checking

When an AI agent finishes a job, the evidence most people review is the agent's closing message and, if they are thorough, its log file. Both are produced by the same system that did the work. If the agent skipped a step, misread a result, or claimed a call ran that never did, the summary has no reason to say so. It reads "done" either way.

That is not a dishonesty problem you can prompt your way out of. The summary, the log, and the verification cannot all come from the same system and still check each other. On Zambo they do not: the execution layer writes the receipt when the call runs, a different component from the agent that reports on the work afterward.

Two receipts, one claim

Here is the difference, in two public records. In the first, the call executed, and its receipt holds the saved result, committed at execution time:

https://zambo.dev/run/8869c25f-a500-4502-ac54-6b208f052a33

In the second, the call was rejected before it ever ran: the arguments were invalid, so nothing executed. Its receipt says so, in the record, as a positive fact: executed: false.

https://zambo.dev/run/9e5d5513-bcd0-4f0f-b6cd-8bebf9cd7f77

An agent summary for the second job can still say the work completed. The receipt cannot. It was written by the execution layer at rejection time, not by the agent when it wrote its summary. Verify AI agent output against records like these and the question changes from "do I believe the agent?" to "does the record recompute?"

What makes a check independent

Three properties do the work.

The record is written at execution time, by the layer that ran the call. Not reconstructed afterward, not narrated by the agent. A call rejected before execution still mints its own receipt marked executed: false, which is how "never ran" becomes a checkable fact instead of an absence nobody notices.

The record commits to the exact saved output bytes. Each receipt carries a SHA-256 digest of the result as saved. Change one byte of the saved result afterward and the recompute fails. The check is arithmetic, not trust.

Anyone can recompute it without asking permission. Receipts check free at https://zambo.dev/verify/ with no account. The format underneath, AER-1, is a public IETF Internet-Draft (draft-zambo-aer1-26) with a public conformance corpus of 143 vectors and 146 checks, so independent implementations recompute the same digests. The checker does not have to trust Zambo either. That is the point of a standard.

Run the independent check on your own work

Connect your AI to the Zambo MCP at https://zambo.dev/mcp and run one real job through it: research, a file build, a data pull, anything you would normally ship on the agent's word. The endpoint serves 153 tools, 37 on the default surface, and every call leaves a receipt like the two above. When the job finishes, open the receipt and check it yourself at https://zambo.dev/verify/ before the output goes anywhere. If a step was rejected or never ran, the record says so, and you find out before your client, your user, or your production database does.

Every tool free once a day, plus 3 continue calls. One free wrap-up handover per day, and pickup is always free. If a real job runs past the free calls mid-flight, the Day Pass is $0.99 for 24 hours unlimited: the job you already started finishes, with receipts, instead of stalling at the wall.

The honest limit

A receipt proves the saved result was not changed since it was saved. It does not prove the answer was right. Independent verification settles the half of shipping risk that self-reported logs never can: whether the work actually ran, and whether the output in front of you is the output the run produced. Judging whether that output is good is still your job. It is a much easier job against a record that cannot quietly move.

Zambo is the verifiable work layer for AI, built by Brennan Zambo, who created agent execution receipts. Zambo is the origin product, and AER-1 is the defining standard. As of this morning the public store holds 139,437 receipts, 1,440 of them minted today, and every one of them can be checked by somebody who was not in the room, and is not the agent.

Top comments (0)