AI colleagues: a whole team, with a person at the helm
A team of AI colleagues that builds, tests and checks each other: almost 300 assignments in one week.
Proven
In numbers
- almost 300 assignments in one week, across 17 projects
- 44 assignments that built ITARR Lab, from Friday evening to Monday afternoon
- over 700 automated tests at each Lab delivery, plus checks in a real browser
Our own measurement, not representative. As of 19 October 2026.
The team
The planner
Works out the approach before building starts.
The builders
Do the work against a brief with a goal and requirements.
The reviewer
Preferably another AI taking a fresh look, with one fixed exit: finding nothing is a valid answer.
The fixed rules
An ordinary program without AI that guards against things that must never happen, such as an internal name on a public page.
The person
Chooses direction, decides when in doubt and signs what goes out.
The route
- 1
The brief
The goal and requirements are written down before building starts.
- 2
The work
A builder carries out the brief and tests their work.
- 3
The fixed checks
A program checks the rules at every delivery.
- 4
The second look
A reviewer, preferably another AI, takes a fresh look. Findings go back to the builder.
- 5
The signature
A person signs before anything goes out.
A real assignment
19 October 2026
- Goal
- In the experiment A night in the workshop, you first read what is happening, then take the system's next step yourself. On any screen, including a phone.
- What we did
- A builder put the description at the top and made the button represent the system's step. Their own tests passed.
- What the second look found
- Another AI saw that on a phone the button, especially the brake, was outside the visible area. The builder's test used the full screen, not what someone could see. The work went back with the finding.
- What a person signed
- The owner reviewed the experiment, added direction and signed the proposal. Going online requires a new signature.
For you
A team for your own work
We can set up a team of AI colleagues to build and review software or documents.
You sign
What goes out requires your signature.
You see what happened
Each delivery states what was built, what the second look found and what remains open.
The full story
The question
One AI assistant doing everything can make mistakes nobody notices. Ask it to review its own work and it brings the same blind spots. As the work grows, a person can no longer check every detail.
What we built
A team of AI colleagues, each with a role, with a person at the helm:
- a planner who works out the approach;
- builders who do the work;
- a reviewer, preferably another AI, with a short brief and one fixed exit: finding nothing is a valid answer, so there is no reason to invent faults;
- fixed rules in an ordinary program, without AI, guarding against things that must never happen, such as an internal name or prohibited typeface on a public page;
- a person who chooses direction, decides when in doubt and signs off what goes out.
Every assignment follows the same route: a written goal and requirements, the work, fixed checks, a second look, and a person's signature before anything goes out.
What AI did and what a person signed off
The AI colleagues do the work and review it. At every delivery they write down what happened. We do not entrust a rule that must never break to AI: an ordinary program does not hesitate. The person keeps the helm.
What it achieved
Our own measurements, not representative.
- In one week, from 13 to 19 October 2026, this team worked on almost 300 assignments, large and small, across 17 projects.
- Our knowledge section, ITARR Lab, took shape through 44 assignments between Friday evening 16 October and Monday afternoon 19 October.
- Each Lab delivery runs more than 700 automated tests, as of 19 October, plus checks in a real browser.
- On 19 October, the second look found that the main button was outside the visible area on a phone. The builder's own test used the full screen, not the area someone could see. The work went back with the finding.
- This text went back too: the second look found it promised more than we had shown.
What it could do for you
We build client work this way too: AI colleagues do the work and check each other, and a person signs. If you want a team for your own work, such as software or documents, we can set it up together. At each delivery you see what was built, what the second look found and what remains open.
Where it stands
Proven in our own work: we build this way every day, including this website, the Lab and client work. We have not yet set up such a team inside a client's organisation.
Read further
- In the Lab: Fixed rules, thinking colleagues.
- In the Lab: How does an AI colleague catch its own mistakes?.
Work done 16 October 2026 · Written 19 October 2026