team-orchestrator

team-orchestrator

AI agent team orchestration plugin for Claude Code. Intelligently routes tasks to subagents or agent teams, composes teams from a role pool grounded in management theory, and self-evolves through session reflection.

Install

# 1. Add the marketplace
claude plugin marketplace add https://github.com/labrinyang/team-ochestractor

# 2. Install the plugin
claude plugin install team-orchestrator@team-orchestrator-marketplace

For local development:

claude --plugin-dir ./team-ochestractor

Prerequisites

Enable agent teams (required for team patterns):

// settings.json
{
  "env": {
    "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1"
  }
}

Skills

Skill	When to Use
orchestrating-work	Core decision engine: subagent vs agent team routing
research-and-review	Parallel investigation with adversarial challenge
feature-development	Independent module development in parallel
competing-hypotheses	Adversarial debugging with competing theories
cross-layer-collaboration	Frontend + backend + test coordination
session-reflection	End-of-session learning capture to .claude.md

Agents (Role Pool)

Organized by Belbin team role categories:

Thinking:

• architect — System design, interface definition (Plant)
• reviewer — Critical evaluation, spec + quality review (Monitor-Evaluator)
• researcher — Deep investigation, evidence gathering (Specialist)

Action:

• implementer — Code building, follows contracts (Implementer)
• debugger — Hypothesis-driven root cause analysis (Shaper)
• finisher — Edge cases, tests, polish (Completer-Finisher)

People:

• coordinator — Team lead, task delegation, synthesis (Coordinator)
• integrator — Cross-module consistency (Teamworker)
• devil-advocate — Mandatory dissent, assumption challenging (Monitor-Evaluator + Safety)

Management Theory Foundation

Rule	Source
Validate Thinking/Action/People role coverage	Belbin Team Roles
Enforce stage gates (plan before implement)	Tuckman's Group Development
Verify 5 conditions before launch	Hackman's Team Effectiveness
Reduce scope when behind, never add agents	Brooks' Law
Team structure mirrors desired architecture	Conway's Law
Max 7 agents, default 3-4	Two-Pizza Rule
Mandatory devil-advocate review	Psychological Safety
Persist retrospective learnings	Agile Retrospectives
Lead manages max 4 direct reports	Span of Control
Every agent has specialty + general context	T-Shaped Skills

Self-Evolution

The plugin learns across sessions:

1. Session start: Hook reads .claude.md with past lessons
2. During work: Agents follow preventive rules from past sessions
3. Session end: session-reflection skill records new learnings
4. Second offense: Repeated mistakes get escalated with stronger rules
5. Universal lessons: Promoted to ~/.claude/CLAUDE.md when confirmed across projects

Decision Framework

Task → Single focused? → Subagent
     → 2+ independent, no communication? → Parallel subagents
     → Need inter-agent communication? → Agent Team
         → Investigation + challenge → research-and-review
         → Independent modules → feature-development
         → Competing theories → competing-hypotheses
         → Frontend + backend + test → cross-layer-collaboration

Default: Start with subagents (cheaper). Escalate to teams only when communication adds genuine value.

License

MIT