Agent teams
How agents group into specialized teams for complex work.
Solo agents handle most things
A typical feature runs through solo agents in sequence. Planner, database, backend, connector, designer, tester, quality gate. That covers most of what you'll build.
But some tasks benefit from multiple perspectives working in parallel. That's where teams come in.
The planning team
When a feature is large enough to warrant it, three specialists analyze the work simultaneously. One focuses on frontend concerns: what data the page needs, how it flows from server to client, what caching strategy makes sense. Another focuses on backend: database changes, API procedures, background jobs. A third focuses on user experience: what the user sees at each step, how it feels on mobile, what happens when something goes wrong.
A lead agent consolidates their analyses into a single plan. Conflicts get resolved. Redundant work gets merged. The result is a plan that's been stress-tested from three angles before any code gets written.
The design team
This one is unusual. When you want to explore different visual directions, the system can spawn multiple designers working on isolated copies of the codebase. Each one takes a different creative approach to the same spec.
You end up with multiple working implementations you can actually run and compare in the browser. Not mockups. Not screenshots. Running code on different ports. You pick the one you like, and the system merges it into your main codebase.
The testing team
Testing splits into two concerns that run in parallel. One agent tests the API layer directly, calling procedures and verifying they return the right data with the right permissions. Another agent opens a real browser and walks through user flows, clicking buttons and filling forms the way an actual person would.
For features that touch authentication, a third tester focuses specifically on auth flows: session handling, token refresh, permission boundaries, what happens when tokens expire mid-action.
When teams activate
You don't choose this. The orchestrator sizes each task and decides whether it needs solo agents or a team. Small and medium features get solo agents. Larger features get planning teams. Design competitions are triggered by a specific command. Testing teams activate whenever the spec includes user flows to verify.