AI · software production
Software production by role-bounded agents
Architecture, API, development, data, tests, review, deployment: each agent has a bounded role, review is adversarial, deployment stays subject to human validation.
Context
Language models write usable code, but a single agent that does everything produces inconsistent systems that are hard to review. The question is not whether AI can write code, but how to organise its production so that it stays under control.
One role per agent, as in a team
One agent frames the architecture, another writes the API, another the tests, another reviews. Each has a scope, written rules and an expected deliverable. No agent deploys.
Adversarial review
The code produced is reviewed by an agent whose role is to find what is wrong, then checked by the tests. What fails goes back to the author. That is what avoids code that looks correct and is not.
A person validates every release
Deployment is a human decision, every time. The log of each agent’s decisions is kept, which makes it possible to understand why a choice was made and to correct it.
Result
Frequently asked questions
Do the agents replace developers?
No. They do part of the production work; developers frame, review, decide and deploy. The craft shifts towards design and review.
How is code quality guaranteed?
Through bounded roles, tests written by a separate agent, adversarial review and human validation before any release.
Which tools?
Language models accessed through APIs, orchestrated by code I control, with a log of every decision. Model providers are replaceable.
Related services
A comparable case on your side?
Tell me which process costs you time; I get back within 24 business hours with a first read, no obligation.