Can one AI agent check another agent's work for compliance?

Yes as a check, no as an approval. One AI agent can review another agent's outputs against policy, look for missing steps, test whether sources support the conclusions and flag anything that breaches a rule. It should not be the final sign-off on that work. A person decides what the findings mean and whether the checked agent's work stands.

Why use a second agent at all?

Agents doing operational work, such as onboarding customers, drafting letters or preparing payments, produce more output than people can review line by line. A second agent that reads every output against the relevant rules can find problems a sample would miss: a disclosure left out of one letter type, a check skipped when a field was blank, a source cited that does not say what the draft claims.

This is the agent equivalent of a second-line check. It is useful because it increases coverage. It is risky if it is mistaken for approval.

Why can't one agent approve another?

Two agents built on similar models, given similar sources, can make the same mistake for the same reason. If the first agent misread a policy, the second may misread it the same way and pass the error. Their errors are correlated in a way that two independent people's usually are not. That is the main argument in why one agent should not approve another agent's work.

There is also an accountability gap. If a checked output causes harm, someone must answer for it. An approval chain made only of agents ends with nobody.

So the checking agent's role is to find and report. Approval of anything material stays with a person, and the checking agent's findings help that person decide where to look.

What evidence does the checking agent assemble?

• The output under review and the action it led to, if any.
• The rule or procedure the output should follow, quoted from the policy library.
• Step checks: whether each required step happened, from the checked agent's own record.
• Source checks: whether each fact the checked agent relied on is in the cited source, and how reliable that source is.
• Bound checks: whether the checked agent acted within its scope, and whether anything it attempted was refused or held.
• Findings: breaches, missing steps, unsupported claims, with severity under your scale.
• Null results: what it checked and found compliant.

To make the check independent, give the checking agent its own brief, its own access to sources and, where possible, a different method. If both agents read the same summary of the policy, the check proves little. The checking agent should read the policy itself, because a summary of a rule is not the rule; the same point is made in agents for regulatory change: the reading is not the work. Checking procedures against policy in general is covered in how to check procedures against policy with an AI agent.

What does a person decide?

What does a check look like?

Take a lender where one agent drafts arrears letters and a second agent checks them. Over a month, the drafting agent produces 2,000 letters.

The checking agent reads every letter against the arrears policy. It finds that 1,940 contain all required information. Sixty omit the reference to free debt advice. All sixty were for customers on a payment arrangement, where the drafting agent used a shorter template. The checking agent also finds one letter quoting an arrears balance that does not match the ledger, because the drafting agent used a balance from before a payment cleared.

The collections compliance manager confirms the sixty omissions are a breach of policy and logs it. The short template is corrected and the sixty customers are sent the information. The balance error is traced to a timing issue in the drafting agent's sources, which is fixed. Neither decision was the checking agent's to make.

Which controls keep the check meaningful?

• Independent brief and sources. The checking agent reads primary sources, not the checked agent's summaries.
• Separate record. Findings are filed in the checking agent's own record.
• People approve. Nothing the checked agent does becomes approved because the checker found no problem.
• The checker is checked. A person samples clean results to confirm the checker is not missing things.

The general principle is in our guide to AI compliance agents.

Where Pryme fits

In Pryme, a checking agent and the agent it checks are certified separately, with separate briefs and sources. The checker can read the other agent's filed verdicts, including actions that were refused or held, but it has no capability to approve them. Approval remains a held step for a named person.