agent-workflow

name: agent-workflow description: Multi-agent development workflow system. Load when orchestrating development tasks, spawning subagents, or managing workflow phases. allowed-tools: Read, Task, Glob

Agent Workflow System

This skill provides the multi-agent workflow orchestration system.

When to Use

Load when:

• Starting a new development task
• Spawning subagents for parallel work
• Managing workflow phase transitions
• Coordinating multi-agent collaboration

Core Concepts

Workflow Phases

1. REQUIREMENTS - PM gathers and clarifies user needs
2. ARCHITECTURE - Architects design solutions in parallel
3. IMPLEMENTATION - Executor manages parallel implementers
4. VERIFICATION - QA validates the implementation
5. REFLECTION - Capture learnings for evolution

Agent Hierarchy

Tier-1 (can spawn subagents):

• PM - Requirements gathering
• Architect - Technical design
• Roundtable - Design synthesis
• Executor - Implementation orchestration
• QA - System verification

Tier-2 (focused workers):

• Explorer - Fast codebase scouting
• Implementer - Focused code writing
• Verifier - Independent verification
• Tester - Test execution
• Contract Resolver - Handle blocked tasks
• Reflector - Extract learnings
• Evolver - System improvement

Agent Definitions

Tier-1 Agents:

• @.claude/agents/pm.md
• @.claude/agents/architect.md
• @.claude/agents/roundtable.md
• @.claude/agents/executor.md
• @.claude/agents/qa.md

Tier-2 Agents:

• @.claude/agents/explorer.md
• @.claude/agents/implementer.md
• @.claude/agents/verifier.md
• @.claude/agents/tester.md
• @.claude/agents/contract-resolver.md
• @.claude/agents/reflector.md
• @.claude/agents/evolver.md

Spawning Pattern

Custom Agents (Dotagent System)

Use lowercase names matching agent definitions:

Task(
  subagent_type: "explorer",    # Fast codebase scouting
  model: "haiku",
  prompt: "Query: ... Output: ..."
)

Task(
  subagent_type: "implementer", # Focused code writing
  model: "sonnet",
  prompt: "Task: ... Boundaries: ... Output: ..."
)

Task(
  subagent_type: "pm",          # Requirements gathering
  model: "sonnet",
  prompt: "Request: ... Output: memory/reports/demand.json"
)

Built-in Agents (Claude Code)

These are different from custom agents above:

Task(
  subagent_type: "Explore",     # Built-in: quick codebase exploration
  model: "haiku",
  prompt: "Find all API endpoints..."
)

Task(
  subagent_type: "Plan",        # Built-in: planning agent
  model: "sonnet",
  prompt: "Plan the implementation..."
)

Subagent Type Reference

Workflow State

@workflows/main.json @memory/state/phase.json