agent-orchestrator

Agent Orchestrator Plugin

A Claude Code plugin that automates multi-agent workflows. Instead of manually copying prompts between agents, describe a feature and let the Software Architect Lead plan, delegate, and coordinate everything.

What This Solves

Before: You type a prompt to the Software Architect → read the plan → manually copy each agent prompt → paste it → switch agents → repeat for every step.

After: /orchestrate Add OAuth2 auth → the orchestrator discovers your installed agents → drafts a plan → you approve → agents run automatically in sequence (parallel only with your permission) → results synthesized.

Installation

From the marketplace (recommended)

# 1. Add the marketplace
/plugin marketplace add Emerging-Disintegration/agent-orchestrator-marketplace

# 2. Install the plugin
/plugin install agent-orchestrator@agent-orchestrator-marketplace

Local development

claude --plugin-dir /path/to/agent-orchestrator-plugin

Prerequisites

• Agents installed in ~/.claude/agents/ — from agency-agents, custom agents, or any source. The orchestrator discovers them dynamically.
• Claude Code v2.1.32+ (for Agent Teams support)
• tmux (optional, for Agent Teams split-pane mode)

Commands

Command	What it does
/orchestrate [feature]	Plan and execute a multi-agent workflow via subagents (asks whether to run in a new Git worktree before making code changes)
/follow-plan [path or description]	Execute a saved plan by path, pick from saved plans, or describe what you want and skip straight to execution (asks whether to run in a new Git worktree before making code changes)
/setup-teams	Enable Agent Teams for the current project
/init-docs	Create docs/ folder with orchestration reference

Quick Start

# 1. Start Claude Code with the plugin
claude --plugin-dir ./agent-orchestrator-plugin

# 2. Create your reference docs (one-time)
/init-docs

# 3. Orchestrate a feature
/orchestrate Build a dashboard showing real-time options Greeks

# 4. Later, re-run a saved plan
/follow-plan docs/plans/dashboard-feature.md

How It Works

You: /orchestrate [feature description]
         │
         ▼
  Software Architect Lead
  ┌─ Scans ~/.claude/agents/ for available specialists
  ├─ Reads STATUS.md for project context
  ├─ Analyzes the feature scope
  ├─ Matches phases to installed agents
  ├─ Presents plan for your approval
  ├─ Asks whether to execute on a separate git worktree (recommended for code changes)
  └─ Waits for approval
         │
    [You approve]
         │
    [Parallel phases? → asks your permission first]
         │
         ▼
  Executes pipeline:
  Phase 1: Agent A (sequential)
         │ output
         ▼
  Phase 2: Agent B + Agent C (parallel, you approved)
         │ outputs
         ▼
  Phase 3: Agent D (sequential, depends on B+C)
         │
         ▼
  Results synthesized → Offers to save plan for /follow-plan reuse

Key Behaviors

Dynamic agent discovery

The orchestrator scans ~/.claude/agents/ and .claude/agents/ at the start of every run. It reads each agent's YAML frontmatter to understand capabilities. No static catalog needed — it works with whatever you have installed.

Parallel requires permission

The orchestrator will never run phases in parallel without asking first. When it identifies independent phases, it'll present them and ask whether you want parallel (faster, more tokens) or sequential (cheaper, easier to follow).

Plan saving and reuse

After execution, the orchestrator offers to save the plan to docs/plans/. Saved plans can be re-executed with /follow-plan — no re-planning needed. Edit the Markdown to adjust before re-running.

Agent Teams vs Subagent Orchestration

The plugin supports both approaches:

• /orchestrate and /follow-plan use subagents — cheaper, the orchestrator relays between agents. Good for 80% of tasks.
• /setup-teams enables Agent Teams — teammates talk directly to each other. Use for complex collaborative work where agents need to debate or share mid-task discoveries.

See docs/ORCHESTRATION-PATTERNS.md for detailed guidance on when to use which.

Reference Docs

Run /init-docs to generate these in your project:

• docs/AGENT-TEAMS.md — Agent Teams setup and usage
• docs/ORCHESTRATION-PATTERNS.md — Workflow patterns, decision flowchart, and reusable plan guide
• docs/plans/ — Directory for saved plans (starts empty)

Compatibility

• Works with any agents in ~/.claude/agents/ or .claude/agents/ — the orchestrator discovers them dynamically