Team Mode workflow
Coordinate two AI engineers with enforced ownership and review boundaries.
Guides follow the current project source. Check the release notes for your installed version.
Create a Team from GPT Agents
Open GPT Agents in the desktop sidebar and select the approved project. Enter the product goal and one concrete success criterion per line, then create the Team.
Copy the separate Engineer A and Engineer B kickoff prompts into the two intended AI chats. They join once and keep their identities until you choose End Team. Keep the project’s access and change policy appropriate for their task.
Planning
- Both AIs join
- Each posts a concise plan
- Each creates a distinct initial implementation task
- Both confirm the split
Execution
- One primary owner per task
- Task scoped path claims
- Distinct implementation scopes
- Explicit handoff when ownership moves
Review and completion
- Other engineer cross-reviews
- Findings return to the task owner
- Success criteria are verified with evidence
- Complete closes the current request, not the persistent Team
Split the work before editing
A persistent A/B Team does not mean two agents should race through one task. Both engineers join, post plans, create distinct initial implementation tasks, and confirm the split. Only then is parallel implementation unlocked.
Each agent owns at most one active implementation task and must hold task scoped path claims for its edits. Shared browser interaction is serialized through an exclusive resource claim.
Cross-review and evidence
Each engineer implements a meaningful scope. The other engineer reviews it and gives concrete feedback when a defect requires correction. Success criteria are verified with actual test/build/browser evidence.
Completing a request closes only that request’s work cycle; the persistent team remains attached until the human chooses to end it.
Avoid duplicate work
If a task owner is blocked, use a deliberate handoff rather than letting the second agent independently implement the same files. Refresh team status and wait for the other engineer to post their work instead of guessing what they changed.
Read the desktop progress view
The app shows the current work request, workflow stage, engineer status, task owners/reviewers, verified success criteria and recommended action. It intentionally hides the raw live coordination feed from the main Team screen.
Completing the request returns the persistent Team to Ready. Pause is temporary; End Team ends the collaboration. An attached AI can register the next human request into the same Team without new kickoffs.