Several agents coding at once. A director that coordinates them.
You ask for a goal and a director splits it between several agents, each in its own isolated copy of the code. You approve the goal and decide what goes in.
The problem
“An AI agent codes well, but you have to stay on top of it.” “If I launch three at once, they step on each other.” “I don't know what each one is doing.”
It happens to technical teams that already code with AI and want to do it properly: with several projects open, watching agents one by one eats the time you meant to save.
How it is done today
Each agent is launched and watched by hand, one at a time, project by project. Someone splits the work, checks they do not collide, reviews what each one delivers and puts the changes together. That person ends up coordinating full time.
How we solve it
- You ask for a goal in one sentence, for example “put three agents on this issue”.
- A lead per goal breaks it into tasks and splits them between the agents.
- Each agent works in its own isolated copy of the code, so they do not step on each other.
- A director oversees every project at once: it sees each agent's progress and spots the one that has stalled.
- What each agent delivers only goes out through a controlled channel that checks the tests and verifies every publication.
- The lead integrates the work and leaves it ready for a person to review.
Result
This is how we develop every day: several projects at once, with one human approval per goal and without watching agents one by one.
What it takes
Your repositories and your issues
Access to the repositories and to the tool where you track tasks. Work starts from a well-described issue.
Automated tests
The publishing channel requires green tests. Without tests there is no way to know whether an agent finished well.
Project rules, in writing
How work is done in each repository: branches, conventions, what must not be touched. It is the first thing each agent reads.
A person in charge
Whoever approves the goals and decides what gets merged. The orchestrator shows them the progress on a dashboard.
How it is built
Three levels: a director that oversees every project, a lead per goal that splits and integrates, and agents that each carry out one task. Each agent works in its own working copy of the repository, on its own branch.
No agent pushes code or opens a merge request on its own. There is a single publishing channel, which checks the tests and verifies every publication. No agent approves its own work.
Each agent logs its progress and the director reads it: that is how the dashboard shows what each one is doing, and how a stalled agent is spotted and relieved. Passwords never reach the model.
Frequently asked questions
How many approvals are needed?
One per goal to start. After that, merging the change and going to production are always done by a person.
What happens if an agent stalls?
The director sees it, because each agent logs its progress. The task is handed to another agent, which carries on from what the first one saved.
Can the agents publish on their own?
No. Pushing code and opening the merge request only go through a controlled channel that checks the tests. Production always belongs to a person.
Related automations
Code review
Four agents review every change before production. A person decides whether it goes in.
Part of AI agent development