How parallel agents build your app
A lead agent plans the work, helpers build the parts in their own worktrees, and only the lead merges the combined result. Here is what happens step by step.
When you turn on team mode, one request becomes a small team of Claude agents. This post walks through what happens between pressing send and seeing the finished change.
1. The lead reads the request
One lead agent starts first. It reads your request and the project, decides how the work splits into independent parts, and does any shared groundwork itself, such as a data model every part depends on. Doing that first means the helpers start from the same foundation.
2. Helpers start, one per part
The lead starts a worker agent for each part. Every worker gets:
- its own Git worktree, branched from the lead's branch at that moment;
- a clear description of its part and the files it is expected to touch;
- the same checks and permissions as any other task.
Workers run side by side, up to the limits your team has set. They never merge into your main branch. When a worker finishes, it waits for the lead.
3. The lead brings the parts together
The lead waits for its workers. Build Depot then merges each cleanly finished worker into the lead's branch, one at a time, and checks for conflicts before each merge. The lead reviews the combined result, fixes anything that does not fit and runs the tests.
4. One change for you to review
Only the lead merges into main, and as one change. You see the diff, the checks and what each agent cost. With autopilot on, the change merges once its checks pass; with it off, you merge it yourself. Either way you can undo it later.
What this means in practice
| One agent | A lead with helpers | |
|---|---|---|
| Large requests | One long run | Parts run at the same time |
| Context | Fills up with every part | Each agent sees its own part |
| Conflicts | None between agents | Isolated worktrees, merged one by one |
| What you review | One change | One change |
Team mode is not free: several agents use more tokens than one. That is why spend limits apply per team, per person and per project, and why every task shows its cost. For small requests a single agent is usually the better choice, and Build Depot lets you pick either for each request.
Want to see the plan before any of this starts? Read Plan first.