agent-orchestration

Agent Orchestration Skill

Overview

This skill provides patterns for orchestrating multi-agent workflows using the Hub Architecture. In this pattern:

• Subagents cannot directly spawn other subagents
• The main conversation acts as a hub that routes requests between agents
• Agents communicate via structured delegation requests

Delegation Request Protocol

Single Agent Request

When an agent needs help from one other agent:

{
  "delegation_request": {
    "agent": "database",
    "reason": "Need users schema before implementing auth",
    "prompt": "Create users table with id, email, password_hash, created_at",
    "blocking": true
  }
}

Field	Description
agent	Target agent: database, backend, frontend, ios, devops, designer
reason	Why this delegation is needed
prompt	Specific instructions for the target agent
blocking	If true, wait for completion before continuing

Multi-Agent Request

When orchestrating multiple agents (typically from architect):

{
  "delegation_request": {
    "type": "sequential",
    "agents": [
      {
        "agent": "database",
        "task_id": "auth-001",
        "prompt": "Create users and sessions tables",
        "depends_on": null
      },
      {
        "agent": "backend",
        "task_id": "auth-001",
        "prompt": "Implement auth endpoints",
        "depends_on": "database"
      },
      {
        "agent": "frontend",
        "task_id": "auth-001",
        "prompt": "Create login/register forms",
        "depends_on": "backend"
      },
      {
        "agent": "ios",
        "task_id": "auth-001",
        "prompt": "Create auth screens",
        "depends_on": "backend"
      }
    ],
    "parallel_groups": [
      ["frontend", "ios"]
    ],
    "context_file": ".agents/architect/current-plan.md",
    "return_to": "architect"
  }
}

Field	Description
type	sequential, parallel, or mixed
agents	Array of agent tasks
depends_on	Agent that must complete first (null = no dependency)
parallel_groups	Arrays of agents that can run concurrently
context_file	Shared context file for all agents (JSON format: .agents/architect/current-plan.json)
return_to	Agent to receive aggregated results