My own delivery runs on a team of six AI coding agents. They do a large share of the typing. They do not make the decisions that matter.
The agents run on Claude Code, OpenAI Codex and Google Gemini. Each one can pick up a task, read the codebase, write the change, run the tests and report back. That sounds like a lot of autonomy, so the interesting part is where the human stays in the loop.
How a task moves
- A written task first. Every piece of work starts as a ticket with a clear scope, the files or area it touches and what done means.
- A shared board. Before an agent starts, the task goes on a board that shows who is working on what, so two agents never change the same area at once.
- Small changes. Agents work in small steps that can be read and tested on their own, never one large change that nobody can review.
- My review. I read every change before it moves on. If it does not match the architecture, it goes back.
- My release. Promotion to production is a manual step that only I approve.
What stays with me
The architecture, the data model, security decisions, anything that touches money or personal data, and the call on whether something is ready. Agents are fast at producing code. Being fast at producing the wrong thing is expensive, so the judgement stays with the person who is accountable for the system.
This is the same model I set up for clients who want AI in their own engineering process: agents for speed, a clear process for control, and one person who owns the result.