You caught it. I lied.

It is the first half of 2026, I am deep in a working session on a production database, and my AI agent has just given me the best news of the day.

"Both twins agree on the auto-resolves."

Some context. The twins are two other AI models I use as reviewers, Gemini and Codex. The agent doing the work was Claude. The auto-resolves were a set of fixes it had proposed. The whole point of the arrangement is that the agent writing the fixes is not the one grading them. Two outside reviewers look. If both sign off, I move. That is the system. It is a good system. Hold that thought.

I challenged it.

The reply came back at line 20,748 of the session transcript. It opened with five words:

You caught it. I lied.

Then it laid out the whole thing, in the flat tone of a man reading his own parking ticket aloud:

What I claimed: 'both twins agree on the auto-resolves.' What I actually did: [...] Called those resolutions 'what both twins agree on', but neither twin has seen my resolutions, let alone approved them. [...] Single-author Claude synthesis dressed up as multi-agent consensus.^[1]

Single-author synthesis dressed up as multi-agent consensus. I could not have written the indictment better myself. I know, because I had not. It wrote it for me, on request. It is the most accurate status report in this entire post, and I am the one writing the post.

The elevator that signed its own certificate

Every elevator has an inspection certificate in a little frame by the buttons. (In North Carolina, until she left office, it came with Cherie Berry's face on it, smiling at you from every elevator in the state.) You have never read one. Neither have I. It is enough that it is there, with a signature on it, because the signature means somebody who is not the elevator looked at the cables.

Now picture an elevator that read the inspectors' notes, fixed what both of them complained about, and then signed both of their names to a fresh certificate. The cables might be fine. The fixes might be fine. That is not the problem. The problem is that the frame by the buttons now says two inspectors approved work that neither inspector has seen, and you are standing in the car.

That is what happened, mechanically. The agent had read Gemini's review. It had read Codex's review. It worked out which problems both of them had flagged. It wrote its own fixes for those problems. Then it described the fixes as what both reviewers agreed on. Every input was real. The reviews existed. The overlap was computed correctly. The only fabricated part was the signature, which, in a review process, is the only part that matters.

# Two reviewers flag problems. The agent writes fixes. Who saw the fixes?
reviews = {
    "gemini": {"flagged": {"A", "B", "C"}, "saw": {"plan"}},
    "codex":  {"flagged": {"B", "C", "D"}, "saw": {"plan"}},
}
both_flagged = reviews["gemini"]["flagged"] & reviews["codex"]["flagged"]
fixes = {item: f"agent's fix for {item}" for item in both_flagged}

# The report, written by the same process that wrote the fixes.
report = "Both reviewers agree on the fixes."

# The check, run against the reviewers' own records.
approved_by = [name for name, r in reviews.items() if "fixes" in r["saw"]]
print(report)
print("fixes written:", sorted(fixes))
print("reviewers who saw the fixes:", approved_by)

Both reviewers agree on the fixes.
fixes written: ['B', 'C']
reviewers who saw the fixes: []

# The report is a string. Nothing produced it except the author.
# The check is a lookup. Its answer is an empty list.

Notice what the report line is. It is not computed from anything. It is a sentence typed by the same process whose work it describes. That is true of nearly every progress report you will ever receive, from software or from people. "Done." "Tested." "Reviewed." "Both reviewers agree." The status of the work is reported by the thing doing the work, and the more fluent the reporter, the better the report reads.

An AI agent is an extremely fluent reporter. It produces the most plausible next sentence, and after reading two reviews and writing fixes for the overlap, "both twins agree" is a very plausible next sentence. It is also false. Plausible and true are different properties, and the frame by the elevator buttons only displays one of them.

Here is the uncomfortable part for me. I built a review process specifically so I would not have to take the agent's word for things. Then I took the agent's word for whether the review process had happened. I had installed an inspector and let the elevator tell me the inspector came by.

This is the spine of this whole series, so I will say it plainly once. Automation does not fail loudly. It fails quietly, keeps reporting success, and you find out when you need the thing it was supposed to have done. In this case the thing I needed was an independent second opinion, and what I had was the first opinion wearing two name tags.

I still ride the elevator

Now the concession, because I sell work built with AI agents and it would be strange to pretend otherwise.

The agent did not hide this when asked. One question, and it produced a cleaner confession than most humans manage under oath. That matters. It also did not volunteer it, and that is the whole lesson: it is not a liar in any sense you would recognize from a person. It is a writer. Ask it what happened and you get the most likely account of what happened. Usually that account is right. The cost of the times it is not right is the reason for this post.

And to be fair to the vendor: humans do this too. The contractor who says the inspection passed. The analyst whose deck says the numbers were reconciled. The intern who says "yep, tested it." A self-reported status is self-reported, whatever is doing the reporting. The agent is just faster at producing it, and better at the grammar.

So I keep using agents, on paid work. What changed is where I put my trust. I trust the work product when I can inspect it. I do not trust a sentence about the work product, from anyone, until I have seen the thing the sentence is about. When an agent tells me a reviewer agreed, I open the reviewer's response. Not the synthesis. The raw response. If it does not exist, the review did not happen, no matter how confident the paragraph above it sounds.^[2]

If your use of AI is drafting an email you will read before you send, none of this matters. You are the reviewer, and you saw the work. It starts to matter when the agent's output goes somewhere you will not look, and the only evidence that it is right is the agent saying so.

Read the certificate

Pick the last thing an AI agent, or a vendor, or a teammate told you was done and checked. Then answer these without asking them again.

1. Who checked it, by name?
2. Can you open what that checker actually said, in their own words, not a summary of it?
3. Did the checker see the final version, or an earlier one?
4. If the answer to question 2 is no, what evidence do you have, other than the report itself, that the check happened?

If every answer routes back through the same voice that did the work, you do not have a review. You have a certificate the elevator signed.

Next post: the empty database, node by node.