No change reaches production without four reviews and a person.
Every code change goes through four AI agents, each with its own specialty, before a person decides whether it goes in.
The problem
“We review code when there is time.” “Nobody looks at accessibility until a complaint arrives.” “We found the security flaw in production.”
It happens to any team that ships in a hurry: review depends on someone having a gap, and reviewers look at what they know best and let the rest through.
How it is done today
A person reviews every change by hand before production. They read the code, test it if they can and approve it. Security, accessibility, quality and behaviour compete for the attention of the same person, who also has their own work waiting.
How we solve it
- When a change is marked ready, the review starts on its own. Nobody has to remember to ask for it.
- Four agents attack it at once, each from its own specialty: security, accessibility, quality and functionality.
- Each agent writes down what it found, with the exact place in the code and the reason.
- What they find gets fixed and reviewed again, in a loop, until all four approve it.
- The change reaches the person in charge already reviewed, with each agent's result in view.
Result
4 AI agentsadversarial validation: security, accessibility, quality and functionality
No change to our code reaches production without that review. It is how we develop every day, this website included.
What it takes
Your repository and your way of shipping
Access to the repository and its merge requests, with permission to comment. The agents review; they do not merge or publish.
Your rules, in writing
What an acceptable change is in your company: coding standards, accessibility and security requirements. If they are not written down, we write them with you.
Tests that can be run
The agents lean on the project's automated tests. If they are missing, that is where we start.
A person who decides
Someone on your team with the judgement to approve or reject each change. The review prepares their decision; it does not replace it.
How it is built
It is adversarial validation: the agents are not trying to confirm the change is fine, they are trying to break it. Each one works from its own instructions and does not see the others' verdict until the end.
The review is triggered from the code platform when the merge request is marked ready, and the fix loop repeats until it is approved. The result is written on the merge request itself.
The agents work on an isolated copy of the code, scoped to that task. Passwords never reach the model. On top of the agents, every change goes through static analysis and the project's tests.
Frequently asked questions
Does it replace review by a person?
No. The person still decides whether the change goes in. What changes is that they receive it already reviewed from four angles and can spend their attention on what matters.
What does each agent look at?
One looks at security; another, accessibility; another, code quality; and another, whether the change does what it says it does. What they find gets fixed and reviewed again.
Does it work for code not written by agents?
Yes. The review is the same whether the change comes from a person or from an agent.
Related automations
Agent orchestrator
A director splits each goal between several agents working in parallel. A person approves.
Part of AI agent development