A multi-agent orchestration framework built on OpenClaw, leveraging a DNA-driven architecture with 60+ roles, an 11-step pipeline, and self-evolution engines for e-commerce analysis, competitive intelligence, business planning, and more.
OpenClaw Multi-Agent Team Framework is a production-grade multi-agent orchestration system running on the OpenClaw platform, currently at v5.3.1. The framework is built around a DNA gene system defining 9 architectural invariants, including a layered architecture (Collection → Processing → Cross-Validation → Evolution), 5 core protocols, structured role prompt templates, domain red-line constraints, deployment tiers (Lite/Standard/Full), data freshness TTL management, knowledge crystallization extraction, and a 5-star confidence rating system.
Execution follows an 11-step hard-dependency pipeline — from complexity check, pattern matching, experience recall, team formation, through parallel execution, cross-validation, quality gate, fix-and-reverify, synthesis report, to post-task reflection — with no steps skippable. Quality assurance is handled by an independent Review Team (7 roles) and Audit Team (7 roles) that only judge without participating in execution, outputting PASS/CONDITIONAL/FAIL. The self-evolution engine comprises 6 gears: Evolution Ledger, Model Abstraction (4-layer decoupling), Cross-Pattern Learning, Capability Frontier, Knowledge Decay, and Structured Extraction.
The role library covers 60+ roles across 11 categories. The public pattern library includes Pattern A (e-commerce), Pattern C (competitive intelligence), Pattern D (business planning), Pattern E (content matrix), Pattern F (tech assessment), and Pattern G (general), with Patterns B/H/R marked as private. The architecture is supported by three core components — EventBus, Blackboard, and Router — with optional cross-team auto-triggering via a file-system event bus.
The framework runs as an OpenClaw Skill, requiring prior deployment of the OpenClaw platform (Node.js v22+) along with an LLM API key and Brave Search API key, triggered via natural language through messaging channels like Telegram/Discord/Slack. It supports Claude Sonnet 4, GPT-4o, Gemini 2.5 Pro, and can be extended with additional skills via ClawHub.
Caveats: The README footer states Apache 2.0, but the LICENSE file is actually MIT — the LICENSE file takes precedence. The repo has no releases and only 25 commits, so the "battle-tested" claim lacks external validation. The model-agnostic claim (80% of value model-independent) has no provided switch configuration examples or benchmarks.