agent-team-builder

name	agent-team-builder
description	Orchestrate specialized agent teams (Architect, Coder, Tester, Reviewer) using high-fidelity communication protocols. Use when a task requires cross-disciplinary coordination, complex architectural changes, or multi-agent parallel workflows.

🏗️ Agent Team Builder

This skill provides the procedural knowledge to orchestrate specialized agent teams within the Initra ecosystem. It enables a "Supervisor" agent to delegate tasks, manage handovers, and ensure architectural integrity across multi-agent workflows.

📋 Orchestration Checklist

Before starting a multi-agent session, verify the following:

1. Role Identification: Are agent blueprints present? (Check .claude/agents/, .cursor/agents/, .codex/agents/, etc.)
2. Topology Selection: Which pattern fits the task? (Pipeline, Fan-out, or Supervisor)
3. Memory Initialization: Is the session context clear? (Framework, Database, Auth selections)

---

🛠️ Orchestration Patterns & Topologies

1. Swarm Topologies

• Mesh (Peer-to-Peer): Equal peers. Use for cross-functional collaboration.
• Hierarchical (Queen-Worker): Architect leads specialized workers. Use for large-scale features.
• Adaptive (Dynamic): Swarm switches topology based on task complexity.

2. Communication Patterns (Handovers)

• Pipeline: Architect -> Coder -> Tester.
• Fan-out: Supervisor -> [Agent A, Agent B].
• Supervisor: Reviewer <-> Coder.

---

🚦 Execution Procedures (The SPARC Cycle)

Agents MUST follow the SPARC cycle for all handovers:

1. Specification (S): The Supervisor/Architect delivers a precise spec for the sub-task.
2. Pseudocode (P): The receiving agent plans the implementation in natural language.
3. Architecture (A): Verify interface compatibility with other active agents.
4. Refinement (R): Implement with TDD. The Coder writes logic; the Tester writes tests.
5. Completion (C): Reviewer validates. Supervisor merges and captures knowledge.

Step 1: Topology Discovery

If agent blueprints are missing, invoke the Initra Engine logic:

• Reference: src/lib/engine/agent-blueprints.ts
• Command: generateAgentBlueprints(ideTarget, variables)

Step 2: Task Delegation

When delegating to a subagent, provide a SPARC Brief:

1. Objective: Precise goal of the sub-task.
2. Context: Relevant files and landmarks.
3. Guardrails: Role-specific "NEVER/ALWAYS" rules from AGENTS.md.
4. Handoff: Next agent in the pipeline or return to Supervisor.

Step 3: Handover & Verification

• Architect to Coder: Deliver API types and Schema definitions first.
• Coder to Tester: Deliver implemented logic and export names.
• Tester to Reviewer: Deliver test results and coverage reports.

---

⚠️ Anti-Drift Protocols

To prevent "Agent Drift" during long sessions:

• Periodic Sync: Every 5 turns, the Supervisor must re-read AGENTS.md and the Venture Blueprint.
• Landmark Locking: Do not allow multiple agents to edit the same landmark simultaneously.
• Truth Source: AGENTS.md is always the final authority on project conventions.

---

📖 Reference Links

• Initra Architecture
• Multi-Agent Recipes
• Behavioral Rules