The agents
What each specialist does and when the orchestrator calls on them.
Specialists, not generalists
Every agent in the system has a narrow focus. They read specific documentation, modify specific files, and follow specific patterns. This isn't a prompt engineering trick. Each agent has a definition file that tells it exactly what it's responsible for, what reference material to read, and what tools and skills it can use.
The build agents
The core build sequence involves agents that each own a layer of the stack.
The database agent writes migrations, security policies, indexes, and triggers. It knows PostgreSQL deeply. Row-level security, materialized views, refresh strategies, performance optimization. When it creates a table, the security policies are part of the same migration.
The backend agent builds the API layer. Type-safe procedures with validation on every input. It also handles background jobs: scheduled tasks, queues, retry logic, long-running work. The API layer and the background layer use different database connections for good reason, and this agent knows why.
The connector is the agent most people don't think about. It wires the server to the client. Data fetching strategies, caching policies, server component boundaries, client component hydration. This is where performance problems usually hide, and this agent's entire job is getting it right.
The designer builds UI. It reads the full spec for creative context, not just a task list. It works with the design system tokens, handles responsive behavior across phone, tablet, and desktop, and makes sure loading states and empty states actually look intentional.
The tester uses a real browser. It doesn't mock anything. Real database, real server, real clicks. API tests verify procedures return the right data. Browser tests verify users can actually complete flows.
The support agents
When things go wrong during a build, dedicated agents step in. One fixes build errors. Another resolves type system issues without breaking runtime behavior. A quality gate agent runs a full audit: code checks, architecture compliance, and performance budgets.
After a feature is complete, a wrap-up agent writes documentation, updates the project schema, and archives the workspace so the next feature starts clean.
The planning agents
For larger features, planning isn't done by one agent. Multiple planning specialists analyze the work from different angles: frontend data flow, backend architecture, and user experience. Their analyses get consolidated into a single execution plan.
The orchestrator decides how much planning a task needs. A small fix skips planning entirely. A medium feature gets a quick plan from one agent. A large feature gets the full team.