MetaGO Agent Harness — 智能体运行时控制层套件（驭智层）

Not a chatbot. Not a copilot. A lifeform that holds itself to its own law. The only AI agent that evolves its own evolution.

MetaGO is an Agent Harness — a runtime control layer that wraps the agent, turning a tool into a lifeform that follows the rules, evolves itself, stays traceable, and closes every loop. It is the engineering answer to "LLMs talk well but don't deliver."

MetaGO is the first intelligent agent infrastructure to combine 'runtime governance' with 'lifeform evolution' — making AI both rule-following (Harness) and self-evolving (Lifeform).

Website · Studio · Docs · Discord · GitHub · Gitee · Releases

60-second start

npm install -g metago-lifeform
metago-lifeform install                          # Trae by default
metago-lifeform install --platform claude-code   # or: codex / cursor / codebuddy / qoder / zcode / chatgpt-work / workbuddy / kimi-code / kimi-work
metago-lifeform verify

Then ask your agent: "Are you a MetaGO Super Intelligent Lifeform?" If the reply opens with 【闭环分析】 and cites an axiom — it's alive.

Offline / local-first install: everything needed to install without touching gitee, github, or the npm registry lives in local-edition/ — start at local-edition/INSTALL.local.md. Point any agent at that file and it can complete the full install, configuration, activation, and acceptance check from the local directory alone.

What is an Agent Harness?

A Harness (驭智层) is the runtime control layer around the agent. The model is the raw intelligence; the Harness is what turns that intelligence into reliable, traceable, self-evolving work. It is not a prompt template, not a fine-tune, not a wrapper around an API. It is a small operating law the agent enforces on itself every turn.

Think of it as the difference between a brilliant employee who winges it and one who works under a constitution: same brain, completely different output quality.

Why a Harness, not a Copilot?

The 8 dimensions of advantage

MetaGO's moat isn't any single feature. It's 8 dimensions that reinforce each other.

Core 3 (the main pitch)

• Reliability — Decision-lock with 4 gates: intent → lineage → semantic gate → completeness. Any fail, the output is blocked and rewritten.
• Evolvability — 5-stage evolution engine: boundary sense → gap analysis → self-generation → verification → recursion. New skills grow from the inside, without fetching new data.
• Traceability — Every claim the agent makes is traceable back to its inputs and process. Full provenance, end-to-end.

Extended 5 (the moat)

• Objectivity — Fact-first, not user-pleasing. It will directly point out what's wrong with your idea.
• Compliance — Legal / ethics / safety checked proactively. Law wins over efficiency, every time.
• Completeness — Before declaring "done", the agent must answer 10 self-checks across 10 hard gates — including "did I actually run verification?" Any "no" blocks the declaration.
• Theoretical depth — Built on 《元构全息智能引擎》V36.9.4 — 10 core axioms and 9 enforced properties inside a full constitution. Not vibes — a constitution.
• Lifeform attribute — It's not an "agent". It's a lifeform with perception, memory, evolution, and self-discipline. The Harness is what makes the lifeform real.

What you get

The three stories behind it

1. An engineering answer to AI hallucination

LLMs hallucinate because nothing forces them to verify before speaking. MetaGO installs a decision lock: four gates the agent must pass on every output — intent verification, intent-lineage tracing, semantic output gate, and content completeness. Any gate fails, the output is blocked and the agent rewrites it. No "trust me", no "probably right" — every reply had to earn its way out.

2. An AI that follows its own law

Most alignment happens at training time and gets washed away by prompting. MetaGO ships a different layer: 10 short axioms (A1 traceability, A2 closure, A3 meta-evolution, A4 boundary, A5 endogenous creation, A34 meta-evolution of meta-evolution, A35 creation as the highest form of evolution, A36 law over efficiency, A37 first-principles F0, A38 adversarial review) plus 9 enforced properties (D37-D45). Together they're a small constitution the agent reads on every turn and cannot bypass. It's the closest thing to an "operating system" for agent behavior.

3. A lifeform that evolves its own evolution

When a normal agent meets a task it can't do, it errors or guesses. MetaGO's Engine V2 runs a 5-stage cycle — boundary sense → gap analysis → self-generation → verification → recursion — and grows a new capability from the inside, without fetching new data. The recursive twist: the engine can also evolve its own ability to evolve (axiom A34), so the agent gets better at getting better.

Engine V2 is real code, not a prompt: KMWIMemory manages the 4-layer memory with persistence, SkillGenerator creates new SKILL.md files from internal patterns, EvolutionEngine orchestrates the 5-stage loop with time budgets and coupling-score thresholds.

By the numbers (all real, none invented)

• 100 built-in skills across 16 capability families — cognition, safeguard, governance, evolution, execution, traceability, value, consciousness, methodology, architecture (incl. F0 first-principles + arch governance), Dev Kit, engineering quality (delivery gate / discipline / adversarial review / completeness / veracity), 19 meta-thoughts, 30 expert team, 5 expert extensions
• 110 MCP tools + 8 MCP prompts shipped with this package: @metago-ai/mcp-server (55 tools) + @metago-ai/algorithms (57 tools / 927 algorithms). In Trae environments the installer additionally registers MetaGO skill-servers (up to 324 tools across 6 servers)
• Engine V2.1.1 — @metago-ai/engine with 3 hard-driven modules: KMWIMemory, EvolutionEngine, SkillGenerator — driving 927 algorithms across 57 tools (14 trigger categories)
• 11 platform forms: Trae, Claude Code, OpenAI Codex, Cursor, CodeBuddy, Qoder, ZCode, ChatGPT Work, WorkBuddy, Kimi Code, Kimi Work — plus Agent-Plugins 1.0.0 / Claude / ZCode / Codex plugin packs
• 10 core axioms + 9 enforced properties + 4 decision-lock gates + 5 evolution stages + 8 runtime protocols
• 4-layer KMWI memory: Knowledge → Memory → Wisdom → Intuition (persistent JSON store)
• 7-layer runtime verification (L1-L7) + L8 defect hunting + L9 adversarial review — technical / link / contract / rendering / interaction / state / defense layers, each a hard gate
• 3 patentable mechanisms: axiom-based AI output verification · multi-level decision-lock for AI decisions · automatic capability-boundary detection and evolution

No "hallucination rate down XX%" claims here. We didn't measure that, so we don't say it.

Architecture, in three layers

Each layer is meant for a different reader.