Documentation / The agent
Agent team
For large jobs, a lead agent distributes tasks among parallel sub-agents, integrates the result, and has it reviewed.
For jobs with several independent parts, the lead agent splits the work into tasks and distributes them among sub-agents working in parallel, each in an isolated copy of the project. At the end, it integrates, tests, and reviews the result.
The team works with Yuli's engine and the provider you configured in Platform mode. It needs Git installed to isolate and merge the work.
How it works
- Plan. The lead agent defines the contracts between layers and publishes a task graph with one role per task.
- Parallel work. Each sub-agent works in its own isolated copy (a Git worktree), on its own branch.
- Integration. Finished tasks are merged into a mission branch. Yuli resolves conflicts and retries failed tasks, with guidance.
- Verification. It looks for conflict markers and runs the project's tests on the integrated result.
- Cross review. A reviewer agent reviews the whole and fixes what it finds.
- Delivery. The result is left as a reviewable task.
Your working copy isn't touched until you approve.
Roles
| Role | Takes care of |
|---|---|
| Backend | Services, controllers, and business logic. |
| Frontend | Interface and components. |
| Database | Schema, migrations, and queries. |
| DevOps | Build, containers, and pipelines. |
| Docs | System documents. |
| Full stack | Tasks that span layers. |
| QA & testing | Strategy and tests. |
| Integrator | Merges and resolves conflicts. |
| Reviewer | The final cross review. |
Configure the team
In the chat header, More actions → Agent team.
| Setting | Default | Range |
|---|---|---|
| Agent team | On | — |
| Sub-agents in parallel | 4 | 1 to 8 |
| Automatic activation | On | — |
| Independent task threshold | 3 | 1 to 20 |
| Run tests on integration | On | — |
| Cross review | On | — |
| Attempts per task | 2 | 1 to 5 |
| Budget per task | US$ 8 | 0 = no limit |
| Budget per mission | US$ 60 | 0 = no limit |
With Automatic activation, the lead forms a team only when the job has at least the threshold number of independent tasks. Otherwise, it works alone.
Budgets
Budgets use the estimated cost. When the limit is reached, the work pauses without losing what's been done. The run resumes after an IDE restart.
Follow the work
- The capsule in the title shows how many agents are working and how many tasks are finished.
- The workboard, in the chat, shows each sub-agent and its task. View full work opens the Team tab.
Team states
Planned, Working, Integrating, Cross review, Paused, Finished, With problems, and Cancelled.
Measurement applies too
The Measurement phase runs on the integrated code, and whatever it finds is passed to the reviewer as one more review check.
With your subscription agent
Claude Code, Kiro, Codex, or Antigravity can launch and follow Yuli's team with the yuli_team_* tools. The team uses the Platform mode provider, not your subscription.
In unattended mode
If the review passes, the team's result is applied directly to your working copy, without a commit. See Unattended mode.