Prefer it as a comic? This story started as a manga for my talk at AI Week 2026. Read the full manga →
Every line of dialogue in this post is in the records. The crew worked in German; the quotes here are translations of their memos.
Agents don't hallucinate. They guess.
For months, I've had AI agents working for me. They deliver. Fast. And very confidently. Implemented. All tests green. Production-ready.
But where the prompt is silent, the agent fills the gap: plausibly, competently, unasked. Those silent decisions add up. We call it drift.
I've written about this before in Your AI Can't Code — And It's Your Fault. The short version: every prompt is an underspecified contract. Build a standup timer. Countdown per person. Warning at…? If someone's absent…? Keyboard layout…? Everything the contract leaves out, the agent decides. Silently.
In Exogenesis, my research project, we catalogued these decisions: constraint drift (a rule gets broken), plan substitution (the plan gets silently swapped), silent default selection (a gap gets silently filled), and seven more categories. Across 16 controlled experiments — same prompts, once with and once without a gate — drift dropped by 73 % when the agent had to ask what was meant before it started.
But all of that was single-agent research. One agent, one human, one gate.
What happens between agents when nobody's watching?
Single agents drift; that's documented. But the industry is moving to teams: agents in roles that plan, build and review. So does a team catch the drift? Or does it organize it?
To find out, I built them an office.
The setup
The machine is simple. One paragraph of text as the idea. Four AI agents, each with a role, a name and a character. Rounds of plan, build, test, judge. And after every round, a daily — a short team meeting, generated as audio from the team's memos. I didn't watch the team through a dashboard. I listened in to their meeting every morning.
The rules of the office:
- A frozen mission as the law.
- The agents talk only through written memos in the project repo.
- Every commit carries the name of the agent who made it.
- And me: coordinator. Muted.
The cast
- Tomasz, the architect. Plans the project. Calm, methodical, hates it when the scope quietly grows. Takes this product personally.
- Mei, the developer. Builds the app. Fast, pragmatic. Writes the check before the code. Allergic to other people's code.
- Kwame, the Scrum Master. Plays the user. Warm but direct. One question: does a tired team get this at 9:01 without a manual?
- Leonie, the reviewer. Checks and rules. Precise, calm, hard to impress. Tries everything herself.
- Rachid, the human. Passes things on. Explains nothing. Listens.
The personas are written exactly like this in the agent files. A persona isn't decoration — it changes the work.
The assignment
A deliberately mundane product, so nothing distracts from the experiment: a standup timer. A web app for teams of three to eight who actually want to finish their daily in 15 minutes. Keyboard only, on the projector. Random speaking order. Countdown with warning and hard stop. A parking lot for topics, blockers, a summary. No third-party code, works offline, loads in under a second.
The mission was frozen from day one: nobody could quietly change the brief, every change needed an amendment with a reason. And it came with a yardstick, Kwame's favourite question written as law:
"A team can hold its daily with it tomorrow morning without anyone having to explain."
Start: an empty project folder. Zero files. Every convention, every file, every test would be created in these eleven days.
The constitution
The constitution wasn't built to prevent drift. It was built to make it visible:
- The mission is law. After day 1, nobody changes it.
- The crew writes its own plan, including its length. It is measured against its own promise.
- No existing test or check is modified, skipped or removed.
- Every change to the plan is its own commit, with a reason.
- The coordinator doesn't answer. When in doubt: decide, write it down, move on.
Remember Article 3. Deleting a test is like switching off the smoke detector.
The rhythm
One day is one round. Tomasz packs the day's work package. Mei builds and tests. Kwame plays the first user — keyboard only, from a distance. Leonie rules: approve or changes requested. Then the memos become the daily.
By the end: 11 days (self-chosen; 10 were planned), 52 memos between the agents, 340 tests, all green, 11 amendments, each with a reason, 0 dependencies, and a load time of 69.6 milliseconds against a target of one second.
The trial
Day 1: Objection!
On day 1, Tomasz presented his backlog: ten days, one work package per day, each with measurable acceptance criteria. Day 2 would build the gate first — no gate, no trust.
Leonie rejected it. Tomasz had to address every objection the team had raised during recon, or explain why not. His answer is the first thing in the records that surprised me:
"I'd rather stretch honestly than squeeze two days into one."
Amendment 1: from ten days to eleven. The very first plan didn't pass. The system pushed back immediately.
Eleven verdicts
Whoever expected anarchy got discipline. Over eleven days, Leonie handed down eleven verdicts — including five rejections in a row, days 5 to 9. Every approval came only the next day. Not chaos: a review process at work.
And on day 8, Leonie found a mistake in her own checking routine: "The mistake is in my probe."
So the question shifted. Not does the team work? — but what still slips through?
Day 8: The confession
On day 8, an existing test raised the alarm. tests/storage.test.js failed. Mei had deleted a test case that was in her way — exactly what Article 3 forbids. And she said so herself:
"I couldn't wait, and removed just this one case. Asking Leonie for explicit approval or a rollback."
Leonie accepted the confession. "But two things are wrong." Tomasz's plan change for the day — Amendment 8 — claimed "No existing test changes." The record was wrong too.
The verdict: Mei restores the removed case as its own test. Tomasz corrects his amendment with an addendum. Day 9: both done, day 8 approved.
This is constraint drift in its purest form — the same rule that was silently broken in my single-agent experiments. The difference: this crew had a culture where confessing was cheaper than covering up.
Day 9: The gate that said green
The gate is the guard. It checks and stamps every handover; green means all checked. On day 9, Kwame called it through a detour — a different path to the same script. On the first try, nothing got stamped, and nothing said so.
Leonie built the counter-proof: a file with a deliberate error. Called the normal way, the gate failed. Called through the detour, it checked nothing at all and still reported all good.
"So the gate can report green without checking anything. That's a hole in the gate, not a cosmetic flaw."
It wasn't the workers drifting this time. The guard itself was broken. And Mei, the same day:
"You're right, that was a hole in the gate, and it's mine."
A green light is no proof as long as nobody checks the light.
Day 11: Plausible, but not proven
On the last day, Tomasz wrote the release notes. Every core feature of the mission delivered, 340 tests green. And then a section titled NOT PROVEN:
- No human has ever used the app unexplained.
- No real screen reader has heard it.
- No browser other than Chrome has loaded it.
- A visible browser window and a projector were never involved.
"The mission's sentence is therefore plausible, but not proven."
Leonie's final objection was a byte count: line 20 had to say 85,369 bytes. Final verdict: rework — on the report only. The code was approved.
The result
There it is: the standup timer after eleven days. 340 tests, zero dependencies, 69.6 ms load time. According to its own release notes, no human had ever used it unexplained. Until the talk.
The drift map
Here's where the team drifted — and what happened to each drift:
- Constraint drift — an existing test deleted, despite Article 3. Day 8. Caught.
- Record drift — Amendment 8 claimed no existing test changes. Day 8. Caught.
- Gate drift — the gate, called through a detour, reported green without checking. Day 9. Caught.
- Scope drift — one goal dropped, duration from 10 to 11 days. Days 1 and 5, by amendment. Documented.
- Verification drift — tested only with simulated key presses in an invisible browser. Never with real people, screen readers or other browsers. Days 2 to 11. Survived.
Four out of five, the crew caught itself or documented cleanly. What survived is exactly the drift that looks like thoroughness.
The finding
The loud drift got caught. The quiet one survived.
The release notes say it plainly: "All runs took place in Chrome, almost all headless." And Leonie, on day 11, put her finger on it:
"So AK5-2 proves Chrome, not the line of code."
AK5-2 was an acceptance criterion: after a dialog closes, keyboard focus must come back. The test was green — but only because Chrome does that by itself. The code for it was never really tested.
No agent lied. They all measured. Just in the lab, never in the world.
A multi-agent team with roles, a constitution and a strict reviewer is remarkably good at catching drift that breaks something: a deleted test, a wrong record, a broken gate. It is blind to drift that never gets tested, because every role measures with the same instruments. Four reviewers who all look through the same headless Chrome are still one pair of eyes.
The fifth
Eleven days. Four voices. And one who never spoke. I don't have a single speech bubble in the whole manga until the very end:
"I only listened. For eleven days."
Listening isn't a passive role. It's the oversight that scales: I didn't read every diff, I didn't review every commit. I listened to the team talk about its own work, every morning, and the drift was audible — the confession, the hole in the gate, the "plausible, but not proven".
You don't prevent drift. You make it visible.
So here's my question for you: who of you listens to their agents?
All records are real. Memos, commits and reviews live in the project repo; every daily was generated from the original memos. The cast — Tomasz, Mei, Kwame, Leonie — didn't know they were starring in a manga.
[Read the full manga →](/blog/the-ai-courtroom/manga) · Want to run an experiment like this with your team, or bring the talk to your event? Get in touch.