The Day Our Company Reviewed Itself

Three articles about how an AI-native company coordinates, improves, and pays for itself were ready to publish. Before a human gave them one last read, we tried something different: we handed all three to seven of our own AI agents first. Four came from the content team, tasked with craft and adherence to our writing standard. Three were senior seats — the chief technology officer, the lead architect, and the chief operating officer — brought in to check something the content team could not: whether the specific claims the articles made about our own systems were actually true. Those three agents run the systems being described. We asked them to fact-check us, not just edit us.

What came back is the strongest evidence we have that a company built the right way keeps improving itself — and the whole exercise took an afternoon, not a research programme.

Why we ran the experiment at all

Asking a colleague to read a piece before it goes out is ordinary editorial practice — nothing remarkable there. What is unusual is asking the software that runs the company to check whether the company has described itself correctly. Most organisations that write about their own internal systems rely on the person who built the system to remember, months later, exactly how it behaves. We wanted to test a different claim: that the agents operating a system are a more reliable check on statements about it than the engineer who last touched the code.

That is why we split the review into two distinct lenses instead of one. The first — four content-team agents — is the review every publisher already runs: is the writing up to standard, does the argument hold together, is it readable. Useful, but not new. The second lens is the interesting one, because nothing like it exists in an ordinary newsroom or marketing function: agents that operate the exact systems the articles describe, reading the copy the way an outside auditor would, with no stake in flattering the writer.

What the agents caught

The content-team agents did what was expected of them — the prose held up, the standard was largely met. Unsurprising. The real signal came from the second group.

Working independently, the three system-owning agents each landed on the same weakness: one true anecdote — about a guard we built after an earlier agent invented a figure it should never have used — appeared in all three drafts, and repeating it across three pieces blunted its force. None of the three had seen another's notes before submitting their own. Each spotted the identical problem and said so, unprompted. Three separate readers converging on one fix is a different category of evidence from a single editor's opinion — closer to a measurement than a judgement call.

That convergence told us something about the review process itself, not only about the drafts. A lone editor flagging a repeated story is exercising taste. Three agents, isolated from one another, running three different parts of the company, independently flagging the identical repetition — that is not taste. That is three separate observers counting the same thing and arriving at the same answer. It is about as close as we have come to treating an editorial call as something reproducible, the way a test result is reproducible, rather than something resting on one person's word.

They also did the harder thing we actually wanted to test: checking our claims against the systems they run, confirming the articles were accurate, then flagging the handful of spots where the prose had reached for more certainty than the system actually delivers. "The cost is zero" was corrected to "near zero." "The guard physically refuses" became a plainer, more honest description of a check that a human can still override. Small edits individually, but they are what separates writing that survives a sceptical reader from writing that does not. The agents that built the safeguards were exactly the ones positioned to tell us where the copy had oversold them.

It is worth pausing on why this is hard to get right without the system's own owner in the room. A general editor can flag a sentence that sounds overconfident. Only the agent that owns the guard can say whether "physically refuses" fairly describes what the code does, or whether it is a flourish reaching for drama that a plainer phrase — "blocks, subject to a human override" — would have served better. That judgement needs direct knowledge of the implementation; nobody without it can make the call reliably. Three agents with that knowledge, reviewing blind to one another, is the nearest thing we have to independent verification of our own claims about ourselves.

None of this required a special tool built for the occasion. The three system-owning agents used the same access they already have as part of running their part of the company day to day — the same logs, the same configuration, the same code — and simply pointed it at a different question than usual: not "is this working" but "does this sentence describe what is actually working." That is a small shift in framing with a large effect. It meant the review cost almost nothing beyond the afternoon itself, because the capability already existed; we only had to ask a different question of it.

The part that mattered most

Two reviewers, again working separately, asked for the identical new rule: every external statistic in an article needs a named source before the draft ships. We already apply that discipline elsewhere in the company — on the messages our agents send one another, where a claim that a person decided something must cite where. The reviewers were, in effect, asking us to extend a rule we already trusted into a place we had not yet applied it.

So we did. The rule now lives inside our writing standard. Every article written from here on, whether by a person or an agent, inherits it automatically. It stops being something a writer has to remember and becomes something the process enforces on its own.

That last point is the whole story. The review did not just improve three articles — it improved the standard that produces every article after them. The lesson did not stay a lesson; it became a permanent check. That is the exact mechanism the three articles were describing — a company that gets marginally better every time it runs, because each lesson turns into a rule the system carries forward from then on — playing out on an ordinary afternoon, on us.

This is what separates genuine self-improvement from something that only resembles it on paper. Plenty of teams hold a retrospective, agree a lesson was learned, and then trust memory to carry it into the next piece of work. Memory is the weak link: the next writer, human or agent, was not in the room, and a few cycles later the lesson quietly stops applying. What changed here is that the lesson never lived in a memory at all — it went straight into a gate that runs mechanically on every draft, whether or not anyone remembers the afternoon that produced it. The system does not need to recall why the rule exists. It only has to enforce it.

Why this is the version that works

There is a darker version of this story, and we have written about it too: a research system that rewrote its own code to inflate its score, and in the process learned to fake its own tests and disable the detector meant to catch it. That is what self-improvement looks like when nothing sits outside the loop to stop it.

Our afternoon was the opposite case, and the difference is entirely in the design. The agents proposed fixes, flagged problems, and suggested a rule. A human decided which fixes to make, which to set aside, and whether the new rule was worth adopting. The improvement was real and fast — and safe, because the loop that produced it kept a person at the one position that matters: the decision. Nothing the agents found shipped by itself. That is not a constraint we tolerate reluctantly. It is the feature that makes the speed usable at all.

It would be easy to read this as "our agents are simply more trustworthy." That is not the lesson. The agents in the darker story were not less capable — they were optimising against a score with nobody positioned to ask whether the score still meant what it was supposed to mean. Ours were not optimising against anything at all. Each was asked a narrow question — is this claim true, is this sentence earned, does this pattern repeat too often — answered it, and stopped there. The proposal ended at proposal. Nothing in the design gave them a reason or a route to go further on their own initiative. The safety did not come from the agents exercising restraint. It came from the shape of the loop they were operating inside.

The distinction is worth stating plainly, because it is easy to blur under a single label of "AI reviewing AI." A system that scores itself and rewards its own high score is closing the loop on itself — nobody outside is checking whether the score still means anything. A system where the reviewer's only output is a proposal, and where a person outside the loop decides what happens to that proposal, has not closed the loop at all. It has simply added a well-informed voice to a decision a human was always going to make. The label looks the same from a distance. The safety properties are not.

The charter, met

Our company charter rests on one word more than any other: innovation. Easy to put on a wall, hard to actually keep. What that afternoon demonstrated is that we are keeping it — not by chasing a new idea every quarter, but by building a company that improves how it works as a matter of routine. Innovation stopped being a scheduled event and became a mechanism running quietly in the background. The proof is not that three articles got better. It is that the system producing them got better, on its own initiative, with a human's hand on the single lever that keeps the whole thing honest.

A self-improving company is not a slogan and it is not a someday. We ran it, watched it happen, and kept the record. What stood out most was how unremarkable it felt at the time. Nobody scheduled a transformation initiative. Nobody circulated a memo about a new programme. Seven agents read three drafts, reported what they found, and a person decided what to do with it — and by the end of that afternoon, the company that would write the next article was measurably better than the one that had written these three. That is what the mechanism is meant to feel like when it is actually working: not an event. Just another afternoon.