A Claude Code skill that orchestrates multiple AI agents from agency-agents (167 specialized agents) to collaborate on any task using tmux for visible parallel execution.

You describe what you need. The skill picks the right 3-8 agents, groups them into phased waves (Research → Build → QA), and launches them in tmux panes where you can watch all agents working simultaneously.

/agency Build a dashboard with appointment scheduling

1. Analyzes your task and selects the best agents from 167 specialists
2. Groups them into waves: Research → Build → QA
3. Shows the team plan for your approval (swap/add/remove agents)
4. Launches tmux session with visible panes per agent
5. Context flows between waves — each wave gets synthesized output from the previous
6. Final synthesis combines all deliverables into .agency/FINAL.md

==== AGENCY ==========================================
TASK: Build a dashboard with appointment scheduling

Selected 7 agents from 167

TEAM PLAN:
+---------------------------------------------+
| WAVE 1 -- Research                          |
|  * UX Researcher -- user flow analysis      |
|  * UX Architect -- component architecture   |
|  * Database Optimizer -- schema design      |
+---------------------------------------------+
| WAVE 2 -- Build                             |
|  * Frontend Developer -- React/Next.js UI   |
|  * Backend Architect -- API endpoints       |
+---------------------------------------------+
| WAVE 3 -- QA                                |
|  * Evidence Collector -- visual verification|
|  * Reality Checker -- final quality gate    |
+---------------------------------------------+

Parallel: 3 | Workspace: /path/to/project

Proceed? [Y / swap / add / remove / parallel N]

• Claude Code installed
• agency-agents installed at ~/.claude/agents/agency-agents/
• tmux installed
• (Optional) Superpowers plugin for enhanced agent discipline

# Clone this repo
git clone https://github.com/mingrath/agency-orchestrator.git

# Copy to your Claude Code skills directory
cp -r agency-orchestrator ~/.claude/skills/agency

git clone https://github.com/mingrath/agency-orchestrator.git ~/.claude/skills/agency

# If you don't have agency-agents installed yet:
git clone https://github.com/msitarzewski/agency-agents.git ~/.claude/agents/agency-agents

In any Claude Code session:

# Basic usage — describe your task
/agency Build a REST API for user management

# With more parallelism (default is 3)
/agency --parallel 5 Create an MCP server for database access

# Marketing tasks work too
/agency TikTok campaign for a pet insurance startup

# Game dev, blockchain, anything — 167 agents cover it all
/agency Build a Roblox experience with DataStore progression

Before launch, you can:

• Y — proceed with the suggested team
• swap UX Researcher for Trend Researcher — replace an agent
• add Content Creator to wave 2 — add an agent
• remove Reality Checker — remove an agent
• parallel 5 — change concurrency
• cancel — abort

Agents don't all run at once. They execute in coordinated waves:

Each wave's output is synthesized into a CONTEXT.md that feeds the next wave, so later agents build on earlier work.

+----------------------------+----------------------------+
|  Coordinator               |  Agent 1                   |
|  (status + synthesis)      |  (working...)              |
+----------------------------+----------------------------+
|  Agent 2                   |  Agent 3                   |
|  (working...)              |  (working...)              |
+----------------------------+----------------------------+

your-project/
├── .agency/
│   ├── task.md              # Original task
│   ├── team.md              # Selected agents + wave plan
│   ├── wave-1/
│   │   ├── agent-name.md    # Agent output
│   │   └── CONTEXT.md       # Synthesized for Wave 2
│   ├── wave-2/
│   │   ├── agent-name.md
│   │   └── CONTEXT.md
│   ├── wave-3/
│   │   ├── agent-name.md
│   │   └── CONTEXT.md
│   └── FINAL.md             # Combined deliverables
└── [project files created by agents]

If you have the Superpowers plugin installed, each agent automatically follows enhanced discipline:

The golden rule: No agent can say "AGENCY_DONE" without proving their work is done.

All 167 agents from agency-agents are available, organized by division:

See references/agent-catalog.md for the full catalog with wave roles and keywords.

The skill scores each of the 167 agents against your task description using:

1. Keyword matching — task words vs agent keywords and description
2. Tech stack context — reads your project files (package.json, etc.)
3. Relevance weighting — exact domain match scores higher than tangential

Then assigns each to a wave based on their role (Research/Build/QA) and presents the team for your approval.

• Agent personas from agency-agents by @msitarzewski
• Built for Claude Code by Anthropic
• Superpowers integration via claude-code-superpowers

MIT