Coautra · Part 2
How does an AI agent that works for you work? Three rules
What rules does an AI need that works for you? The same as a good team. You just have to write them down once more for machines — and then a program enforces them, every day, without exception.
The three rules come from a platform that AI agents build and that we laid open in July. Further down: how far they have carried since.
Three rules for agents that work
1. Check after every step. After every merge the line is locked until the main branch is demonstrably green. That costs minutes and prevents two changes from building on an error of the first.
2. «Half done» is a state of its own. If an agent leaves something out, the task does not start from scratch. It restarts at the implementation with the same working copy and reads the requirement once more. A half-finished task is neither «done» nor «failed» — it has its own path.
3. Whoever checks is not whoever does. A separate agent writes the tests from the acceptance criteria. The agent that implements is given them. The same separation every good team draws between development and testing.
None of these rules is an invention. They are rules from working with people. The difference: a program enforces them, even on the day nobody thinks of them.
The cycle keeps running
In July this blog carried Not a promise. Proof, backed by numbers from the Git system. Proof only counts as long as it keeps running. So: the same query once more, the same three core repositories, as of 5 October 2026.
| 21 July | 5 October | |
|---|---|---|
| Backend (API) | 682 | 877 |
| Web | 669 | 848 |
| Process system | 387 | 545 |
| merged merge requests | 1,738 | 2,270 |
In eleven weeks 532 changes were added, almost 50 a week — our own calculation from the two queries. Each of them reaches the main branch via tests and review. And the process system itself grew by 158 merge requests: the workflow that leads the agents is developed further with the same workflow.
A human at the start and at the end
The start of the chain is a human who says what they need. The end is a human who decides what gets delivered. In between the agents work — in steps anyone can read up on.
What this means for your office
You do not have to build a platform to benefit. The three rules apply to any work you hand to a machine: check after every step, know «half done» as a state, and separate the one who checks from the one who does. In the AI back office the agents work by the same pattern — with mails and documents instead of code.
Which of these three rules is missing from the automation you already have?