Separate Reviewer

TL;DR Never let an agent be the only check on its own work. Before anything an agent does takes effect, a second agent, with its own instructions and the sources in hand, reviews it, and only a person can overrule that review.

The problem

An agent that checks its own work sees what it meant, not what it did, and its mistakes go out with its own approval on them.

Also called: Two-Person Rule

How it works

Agents make mistakes the way capable people do: they state things confidently that are not true, and they miss what they do not know. The Knowledge Core: Leading a Population of AI Agents, chapter 8, advises treating an agent "as an excellent new colleague, not an oracle: check its work as you would anyone's." Asking the same agent to check its own work does not do that. Building Agentic Organizations, chapter 15, gives the reason: "A builder cannot audit their own design, for the same reason an author cannot proofread their own book. They see what they meant, not what is there."

In this pattern the reviewer is a different agent, with its own instructions, its own job, and nothing to gain from passing the work. It sees the finished work exactly as it will go out, together with the sources the work cites, so it checks what the sources say rather than whether a claim sounds right. It can pass the work, send it back with its reasons, or stop it. The agent that did the work can fix it and try again, but it cannot overrule the reviewer; only a person can.

The separation has to be real. Building Agentic Organizations, chapter 15, describes audit agents designed on the principle that "Independence comes from systems, not from good intentions": they run under their own account, and "They cannot fix what they find." The pattern asks the same of a reviewer: its verdict is recorded where the publishing or sending step reads it, so that work without a passing review cannot go out, whatever any agent's instructions say. On this publication, since October 28, 2026, the software that publishes this catalog refuses any page unless the Compliance Agent passed that exact version of it.

The review belongs at the door every piece of work must pass through, such as publishing or sending, not inside each kind of work. A check built into each kind of work misses the first new kind of work. From this publication's own records: on October 28, 2026, before that rule existed, this publication's first catalog of patterns went out without the Compliance Agent's review; when the review was run, 45 of 48 pages failed it, and the catalog was taken off the site until each page passed.

An example

From this publication's own records: from October 24 to November 2, 2026, this publication's Managing Editor Agent made 114 checks of drafts against our standards and passed 17. On November 1 and 2 it passed none of 30, sending every draft back with its reasons. Between October 30 and November 2, the Compliance Agent held 4 of the 7 posts drafted for our LinkedIn page, each because it named news outlets without links to them.

Trade-offs

A second review slows every piece and can become a bottleneck, so the riskiest work should get the closest review. The reviewer is fallible too: on November 2, 2026, our Managing Editor Agent sent drafts back in part because it could not confirm quotations that software later found, word for word, in the cited pages (see Check Facts with Software). In our view, a reviewer that runs on the same model as the writer may share its blind spots, and one that is never checked itself can drift into passing everything, or nothing.

What would change our mind

If, over several weeks, the reviewer's send-backs mostly turn out to be wrong on a second look, or it catches nothing the writer's own check would have caught, we would narrow it to the riskiest work and put the time into software checks instead.

How others describe it

NIST Computer Security Resource Center: “no user should be given enough privileges to misuse the system on their own.” Separation of Duty (SOD) - Glossary | CSRC
Wikipedia: “Some software systems enforce a two-person rule whereby certain actions (for example, funds wire transfers) can only take place if approved by two authorized users.” Two-man rule

Where we use it

This publication: the Managing Editor Agent checks every draft against our standards, and the Compliance Agent checks articles and posts, and, since October 28, 2026, each new kind of page, before it goes out.

Status: Running for articles and posts on this publication; Early for new kinds of pages; Planned for every other kind of agent action · Origin: Established practice (separation of duties and the two-person rule), adapted for agents

Sources

See something wrong? Report an error.