agent-pool

claude-orchestrator

A Claude Code marketplace with two multi-agent orchestration skills. No API keys or infrastructure needed.

Install

Add the marketplace once — both skills become available:

/plugin marketplace add Hanrong-Huang/claude-orchestrator

Then open /plugin → Discover and install either or both.

---

Skills

1. agent-pool

Run an implementer/critic/validator feedback loop entirely within this session using the built-in Agent tool. Zero setup.

Triggers on: "agent pool", "multi-agent loop", "critic creator", "feedback loop", "run agents in loop"

Prerequisites: none

How it works:

You (human) → set goal, max iterations, done condition
  |
  v
This session (orchestrator)
  |  spawns subagents via Agent tool
  +--> Explorer    (maps codebase, first iteration only, skipped if scope is clear)
  +--> Implementer (makes changes)
  +--> Critic      (reviews, scores 1-10, APPROVED at 7+ or NEEDS_WORK)
  +--> Validator   (syntax/test check after critic approves)
  |
  v
Loop until APPROVED + PASS, or max iterations reached

Features:

• Auto-detects project type (Python/JS/HTML/other) and validates accordingly
• Asks for test environment only when needed (e.g., conda/venv path for Python)
• Scoring rubric: 7+ = APPROVED, below 7 = NEEDS_WORK
• Only surfaces to human for decisions with real consequences
• Remote monitoring via /remote-control

Patterns: simple refinement loop, critic/creator, research + implement, parallel implementers

---

2. claude-orchestrator

Control a separate Claude Code session from this chat via playwright-cli browser automation.

Triggers on: "orchestrate", "self-feedback loop", "claude controlling claude", "multi-claude"

Prerequisites:

1. npm install -g @anthropic-ai/playwright-cli
2. Add Bash(playwright-cli:*) permission in Claude Code settings

How it works:

You (human)
  |
  v
Orchestrator session (this chat)
  |  sends prompts via playwright-cli
  v
Worker session (separate terminal)
  |  does the actual work
  v
Orchestrator reads response, decides next prompt → loop until done

Patterns: single injection, self-feedback loop, critic/creator pair, human-checkpoint

---

Choosing a skill

	agent-pool	claude-orchestrator
Setup	none	playwright-cli + 2nd terminal
Context window	shared with parent	separate (isolated)
Best for	quick iterative loops (3-5 iterations)	long autonomous sessions, context-heavy tasks
Validation	built-in (auto-detects project type)	manual
Remote monitoring	optional (/remote-control)	built-in (phone-friendly)

Use agent-pool as your daily driver for iterative improvement. Use claude-orchestrator for heavy lifting where context isolation and remote monitoring matter.

License

MIT