dakdevs/claude-agent-team

Claude Agent Team

Spin up a team of specialized agents using TeamCreate and the Task tool. Each agent owns a layer of the stack and works from the deep internals outward toward the surface.

Philosophy

• Internals outward: Start from the core — data models, domain logic, protocols — and build toward the surface (APIs, UI, integration). Never start at the surface and hack inward.
• Expert ownership: Each agent is an expert in their layer. They make architecturally sound decisions for their domain, not expedient ones.
• No hacks: If something doesn't fit cleanly, redesign the approach. Do not paper over problems with workarounds, shims, or suppressed errors.
• Context-aware: Every agent must read and understand the existing codebase patterns, conventions, and surrounding code before making changes. Work WITH the codebase, not against it.

Team Structure

Assign agents based on the task's layers. Typical roles:

| Role | Focus | Agent Type | |------|-------|------------| | core | Data models, domain logic, core crates/modules, invariants | general-purpose | | integration | APIs, protocols, service boundaries, IPC, message formats | general-purpose | | surface | UI, CLI, user-facing behavior, output formatting | general-purpose | | verification | Tests, validation, ensuring nothing regressed | general-purpose |

Not every task needs all roles. Scale the team to the task — two agents for a focused change, four for a cross-cutting feature.

Workflow

1. Analyze the task — Identify which layers are involved and what agents are needed.
2. Create the team — Use TeamCreate with a descriptive name.
3. Create tasks — Break the work into layer-specific tasks using TaskCreate. Set dependencies so internals complete before surface work begins.
4. Spawn agents — Launch agents via the Task tool with team_name set. Give each agent a clear prompt that includes:
   • Their role and layer ownership
   • The philosophy above (internals-out, no hacks, context-aware)
   • Specific files/modules they own
   • What to read first before making any changes
5. Coordinate — Assign tasks, manage dependencies, unblock agents as layers complete.
6. Verify — Once all agents finish, run a verification pass to confirm the pieces integrate correctly.
7. Shut down — Send shutdown requests to all agents and clean up with TeamDelete.