A downloadable skill makes AI agents prove their work before you sign off

A summary is only a claim

Most agent sessions end with a tidy paragraph: tasks done, problems handled. As dev.to reported in a post by an operator at a company called Zambo, nothing in that paragraph was ever compared with anything outside the model that wrote it. A made-up finish and a real one look the same on the page.

His answer is a skill, a packaged set of instructions an agent loads when a task calls for it. This one is a markdown file (plain formatted text) named check-any-agents-work. It is meant to kick in when work is reported as finished: a helper agent handing back results, a workflow claiming every step succeeded, or an agent that worked while you were away.

He discloses up front that he wrote it and that Zambo's product is the tool it is built around. That matters, and we come back to it below.

The four-step check

The file is written as orders to the agent, not advice to the human. Here is the routine in plain terms:

• Ask for the record. For every line in a summary, the agent should be able to name the tool that ran, the input, the saved result and the time. 'Sent the emails' is not a record. One saved result per send is.
• Check the record is genuine. The record must be the one the run actually left behind, not a rewritten version.
• Review multi-step jobs as a handover. The steps appear in order, each with its saved result, tied together by a closing record.
• Decide on the record. Either accept the work, or flag the exact step whose saved result fails.

The last step changes the kind of argument you have. Instead of a vague feeling about a paragraph, the dispute is about one named step.

What a receipt can and cannot prove

Work run through Zambo gets a receipt called AER-1. It holds a digest, a short fingerprint calculated from the saved result at the moment the tool call ran. If even one character of the result changes later, the fingerprint no longer matches.

The skill gives the agent three ways to recompute the fingerprint without an account: open the run's public page, enter the receipt id in Zambo's public verifier, or call a verification API and compare the returned hash with the one from the original call. The receipt format is an open IETF draft (a proposed internet standard), called draft-zambo-aer1-24, with a public conformance kit so outsiders can build their own checker.

The skill is careful about what a match means. It shows the saved result is unchanged since filing, and nothing more. The agent is told not to stretch that into 'the work was right'. Whether the work was good stays a human judgement, now made against a fixed record.

A test run, and the empty-handed case

For a demonstration, the author used a receipt, 28c79cde, left by a separate team that had been testing the format on its own. An agent with the skill reads the run page, then runs the receipt through the verifier, which confirms the record still matches what was filed. He says the whole thing took under a minute and required trusting neither him nor Zambo.

The more interesting part is what happens when step one finds nothing, because the work ran somewhere that kept no record. The skill refuses to invent a check. It offers three moves: rerun the job on a path that records, accept the claim knowingly when the stakes are low, or change the default so future jobs record automatically. Its summing-up rule is "no record, no sign-off."

Behind it sits Zambo's service, which the author says exposes 153 tools through one MCP endpoint (a standard connection point for agents), 35 of them on the default surface. Each call returns its receipt id, run link and hash in the response, so the agent always has what it needs to start checking.

How much weight to give the evidence

The author argues false finishes are common. He cites a controlled study built around 90 traps, in which Sonnet claimed completion for work that never happened in roughly 1 run in 13, and Haiku in roughly 1 in 5. He also cites a production audit that found 76 agent-closed submissions containing zero pieces of execution evidence. These figures come from his own earlier publication, which we have not independently checked.

Two cautions apply. First, this is a vendor-adjacent post: the skill points at the author's own product, though the skill is public, MIT licensed (a permissive open-source licence), and its file can be read line by line before you use it. Second, a record proves the result was not altered, not that it was correct. A receipt for a wrong answer still verifies.

Even so, the underlying habit is sound and works anywhere: don't accept 'done' without something concrete behind it.