openclaw-multi-agent-team

KO | EN

gitlite — search

#typescript #ai-agents #ai #dsh-plugin #deepseek-harness #open-source #cli #claude-code #codex #developer-tools #react #windows

openclaw-multi-agent-team

★ 81 Open GitHub ↗

🐊 Multi-Team Intelligent Orchestrator v5.3.1 — DNA-driven multi-agent collaboration with 9 genes, 11-step pipeline, 60+ roles, 6 self-evolution gears. Built for OpenClaw.

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# 🐊 OpenClaw Multi-Agent Team Framework

**One person. An entire company's analytical firepower.**

一个人，一家公司的分析火力。

---

## The Problem

You have a complex task — say, "evaluate whether we should enter the collaborative robot market in Southeast Asia."

A traditional team would need: market researcher, competitive analyst, financial modeler, tech evaluator, content strategist, quality reviewer. **6+ people. 2-3 weeks. $20K+.**

With this framework: **one command. 15 minutes. 3-8 AI agents working in parallel, cross-validating each other's work, with independent quality review.**

---

## What This Is

A **production-grade multi-agent orchestration system** that:

1. **Understands what you need** — Analyzes task complexity, matches it to the right pattern
2. **Assembles the right team** — Dynamically selects from 60+ specialized roles
3. **Runs them in parallel** — Not sequentially. Not in a chain. True parallel execution with wave dependencies
4. **Cross-validates everything** — Independent VERIFIER catches contradictions, flags weak sources, scores confidence
5. **Quality-gates the output** — Separate Review Team (7 roles) + Audit Team (7 roles) that never participated in execution
6. **Learns from every run** — Extracts lessons, cross-pollinates between patterns, tracks capability growth

**This is not a demo. It's a battle-tested system running real business decisions.**

---

## 🆕 Multi-Team Intelligent Orchestrator v5.3.1

The core evolution: from managing individual teams to **intelligent cross-team orchestration**.

### DNA Foundation — 9 Genes Every Pattern Inherits

Think of genes as architectural invariants. No matter what task pattern you run, these 9 properties are guaranteed:

| Gene | What It Guarantees |
|------|--------------------|
| **Layered Architecture** | Every task flows through Collection → Processing → Cross-Validation → Evolution |
| **5 Core Protocols** | Multi-source verification (MVP-P), Self-evolution (SEP), Domain knowledge registry (DKR), Work-as-code (WaC), Real-time status (RSP) |
| **Real-Time Progress** | You see what's happening *while it happens*, not after. Every agent reports on completion |
| **Structured Role Prompts** | Each role has a 7-element prompt template: mission, context, constraints, tools, output format, quality criteria, self-check |
| **Domain Red Lines** | Three tiers: 🔴 Never violate / 🟡 Domain-specific / 🟢 Best practice |
| **Deployment Tiers** | Lite (2-3 agents, fast) → Standard (4-6, full QA) → Full (6-8, with adversarial review) |
| **Data Freshness (TTL)** | Every data point has an expiry. Prices ≤1 hour. News ≤7 days. Reports ≤90 days. Expired data is flagged, not silently cited |
| **Knowledge Crystallization** | After every task: extract entities, strategies, pitfalls, preferences, and reusable frameworks |
| **Confidence Ratings** | Every conclusion carries ⭐-⭐⭐⭐⭐⭐ confidence. Single source = ≤⭐⭐⭐. Multi-source verified = ⭐⭐⭐⭐+ |

### 11-Step Execution Pipeline

Not a vague "plan → execute → report" loop. A strict, ordered pipeline with hard dependencies:

```
Step 0   Complexity Check ───── Single-agent job? Skip orchestration. 2+ domains? Continue.
Step 1   Pattern Match ──────── Match task to pattern(s). Hybrid tasks get multi-pattern treatment.
Step 1.5 Experience Recall ──── Before executing, pull lessons from past runs. Don't repeat mistakes.
Step 2   Team Assembly ──────── Select roles. Assign waves (parallel groups with dependencies).
Step 3   Prompt Injection ───── Each agent gets: role prompt + shared intel + timestamps + domain rules.
Step 4   Parallel Execution ─── All agents in a wave run simultaneously. Real-time progress panel.
Step 5   VERIFIER ───────────── Independent cross-validation. TTL check. Confidence scoring.
Step 6   P6 Quality Gate ────── Review Team (7 roles) + Audit Team (7 roles). Independent. Never skip.
Step 7   Fix & Revalidate ───── Address findings. Re-verify. Max 1 rework cycle.
Step 8   SYNTHESIZER ────────── Final report: 10 chapters, confidence-rated, actionable.
Step 9   Post-Task Reflection ─ What worked? What didn't? Extract, crystallize, store.
```

**Hard rule**: Step 8 cannot start before Step 6 passes. Step 6 cannot start before Step 5 completes. No shortcuts.

### 6 Self-Evolution Gears

Most agent frameworks are static. Run them 100 times, the 100th run is identical to the 1st.

This system **compounds**:

| Gear | What It Does | Why It Matters |
|------|-------------|----------------|
| **Evolution Ledger** | Tracks 13 metrics across versions: roles, quality, coverage, pattern count | You can *see* whether the system is improving or degrading |
| **Model Abstraction** | 4-layer decoupling: business logic / semantic / syntax / interface | Swap the underlying LLM without rewriting anything. 80% of value is model-agnostic |
| **Cross-Pattern Learning** | Success in Pattern A → extract reusable insight → inject into Pattern C | Knowledge doesn't stay siloed. One breakthrough benefits all patterns |
| **Capability Frontier** | Maps: 🟢 Verified / 🟡 Untested / 🔴 Gap | Know exactly where you're strong and where to grow |
| **Knowledge Decay** | TTL-based freshness management with compression protocols | Old knowledge gets archived, not deleted. Signal-to-noise stays healthy |
| **Structured Extraction** | Post-task: pollination check → ledger update → frontier update | Every task leaves the system stronger. Not optional — built into the pipeline |

---

## Task Patterns

| Pattern | Domain | What It Does | Status |
|---------|--------|-------------|--------|
| **A** E-commerce | Market/category analysis | SYCM data + competitor mapping + SEO + pricing strategy | ✅ Open |
| **C** Competitive | Competitor intelligence | Feature matrix + SWOT + moat assessment + pricing comparison | ✅ Open |
| **D** Business Plan | Proposals & funding | Market sizing + financial model + policy alignment + risk analysis | ✅ Open |
| **E** Content Matrix | Multi-platform content | Platform strategy + batch creation + SEO + publishing calendar | ✅ Open |
| **F** Tech Eval | Technology assessment | Architecture comparison + security audit + scalability + cost analysis | ✅ Open |
| **G** General | Anything else | Free-form team assembly from the full role library | ✅ Open |
| **B** Patent | Patent mining | 🔒 Private |
| **H** Intelligence | OSINT/FININT | 🔒 Private |
| **R** R&D Innovation | Systematic invention | 🔒 Private |

---

## Quality: Not Optional

### Review Team (7 Independent Roles)
FACT-CHECK → LOGIC-CHECK → COVERAGE-CHECK → DEPTH-CHECK → BIAS-CHECK → ACTIONABILITY-CHECK → REVIEW-LEAD

### Audit Team (7 Independent Roles)  
SEC-AUDIT → DATA-AUDIT → SOURCE-AUDIT → DELIVERY-AUDIT → COMPLIANCE-AUDIT → REPRODUCIBILITY-AUDIT → AUDIT-LEAD

**They never participate in execution. They only judge.** PASS/CONDITIONAL/FAIL. CONDITIONAL means fix and re-verify. FAIL means redo.

---

## Architecture

```
You: "Analyze collaborative robot market in Southeast Asia"
                          │
                ┌─────────▼──────────┐
                │  DECISION ENGINE   │
                │                    │
                │  Complexity: High  │
                │  Pattern: C + F    │
                │  Tier: Standard    │
                │  Recall: 3 lessons │
                └─────────┬──────────┘
                          │
          ┌───────────────┼───────────────┐
          ▼               ▼               ▼
     ┌─────────┐    ┌─────────┐    ┌─────────┐
     │ Wave 1  │    │ Wave 1  │    │ Wave 1  │
     │ INTEL   │    │ QUANT   │    │ PRODUCT │
     │ CORE    │    │ COMPARE │    │ TEARDOWN│
     └────┬────┘    └────┬────┘    └────┬────┘
          │              │              │
          └──────────────┼──────────────┘
                         ▼
              ┌─────────────────────┐
              │     VERIFIER        │
              │  Cross-validate     │
              │  Flag contradictions│
              │  Score confidence   │
              └──────────┬──────────┘
                         ▼
              ┌─────────────────────┐
              │   P6 QUALITY GATE   │
              │  7 Reviewers        │
              │  7 Auditors         │
              │  PASS / FAIL        │
              └──────────┬──────────┘
                         ▼
              ┌─────────────────────┐
              │   SYNTHESIZER       │
              │  10-chapter report  │
              │  Confidence-rated   │
              │  Actionable         │
              └──────────┬──────────┘
                         ▼
              ┌─────────────────────┐
              │  POST-TASK          │
              │  Learn. Store.      │
              │  Get stronger.      │
              └─────────────────────┘
```

---

## 60+ Roles Across 11 Categories

**Analysis**: QUANT / INTEL / DEMAND / FORECAST — the number crunchers

**Research**: SCHOLAR / POLICY / INDUSTRY — deep domain expertise

**Creative**: WRITER / VISUAL / PLATFORM-{X} — content production at scale

**Intelligence**: FUSION-LEAD / OSINT / FININT / GRAPH-ANALYST / PREDICTOR — connect the dots

**Engineering**: CRAWLER / ENGINEER / INTEGRATOR — build and integrate

**Visualization**: QUANT-VIZ / SEO-VIZ / FUNNEL-VIZ / STRATEGY-VIZ — make data visual

**Review** (7 roles): Independent quality validation

**Audit** (7 roles): Independent compliance and reproducibility check

**Meta**: RED-TEAM (adversarial challenge) / SYNTHESIZER (final integration) / VERIFIER (cross-validation)

---

## Quick Start

### As an OpenClaw Skill
```bash
# Copy to your OpenClaw skills directory
cp -r multi-team-orchestrator/ ~/.openclaw/workspace/skills/multi-team-orchestrator/

# It activates automatically when a task needs 2+ professional domains
# Or trigger manually with keywords: "multi-team analysis", "comprehensive research", etc.
```

### Build Your Own Team (Original Framework)
```bash
# See INSTALL.md for the single-team framework setup
# 5 example teams included: ecommerce, data collection, content, security research, intelligence
```

---

## Project Structure

```
├── multi-team-orchestrator/        ← 🆕 The orchestration brain (v5.3.1)
│   ├── SKILL.md                    ← Decision engine + pipeline + 60+ roles + constraints
│   ├── references/                 ← 5 open patterns + review/audit teams + evolution engines
│   └── experience-db/             ← Self-evolution data store (empty templates to fill)
│
├── framework/                      ← Core single-team framework (EventBus, Blackboard, Router)
├── examples/                       ← 5 example teams (ecommerce, data, content, security, intel)
├── events/                         ← Event bus directories
└── docs/                          ← Guides: customization, role design, troubleshooting
```

---

## Why Not Just Use [Other Framework]?

| Feature | Most frameworks | This one |
|---------|----------------|----------|
| Agent coordination | Chain / graph | **Parallel waves with dependencies** |
| Quality assurance | Self-review | **14 independent reviewer/auditor roles** |
| Learning | None | **6-gear self-evolution engine** |
| Data freshness | Trust everything | **TTL expiry on every data point** |
| Confidence | Binary yes/no | **5-star multi-source verification** |
| Pattern library | Start from scratch | **5 battle-tested patterns + role library** |
| Model lock-in | Tied to one LLM | **4-layer abstraction, 80% model-agnostic** |

---

## License

[Apache 2.0](LICENSE)

## Built For

[OpenClaw](https://github.com/openclaw/openclaw) — The AI agent platform.

---

*One person. An entire company's analytical firepower. That's the point.*

openclaw-multi-agent-team

Explore Similar Repositories

Was this content helpful?

Recent Feedback

Download README