01 — CLI
Orca CLI
Provider-neutral coding agent — 9 providers, 41 tools, MCP server, 5 execution modes. One agent surface over every model.
中立于厂商的编码智能体——9 家模型、41 个工具、MCP 服务、5 种执行模式。一个统一的智能体界面，覆盖所有模型。
AI Agent Systems Engineer · Data Intelligence Architect
Maurice Wen
「AI 智能体需要的是基建，而不是提示词。」
I build the orchestration layer — routing, context, multi-model collaboration, and verification — that turns raw LLM capability into reliable enterprise systems.
我构建智能体的编排层——路由、上下文、多模型协作与验证门——把原始的大模型能力，变成企业可以信赖的系统。
我的建立 · Origin
From national-scale data to autonomous agent fleets.
从国家级的数据平台，到自主运转的智能体舰队。
Fifteen years turning messy operational data into systems people trust — then rebuilding that trust one layer lower, beneath the models.
十五年，把杂乱的经营数据变成别人愿意信赖的系统；然后，在模型之下更低的一层，重建这份信赖。
01 / 2010–2023
Data.
02 / 2023–2025
Pivot.
03 / 2025–now
Fleet.
Data at national scale
Senior Product Operations Engineer at Kingdee (金蝶), leading data platforms for China's Ministry of Industry and Information Technology — monitoring 1M+ enterprises, saving ~2M CNY/year in data cost, and running privacy-preserving federated learning with WeBank.
金蝶高级产品运营工程师，为工信部（MIIT）搭建数据平台，监测超过 100 万家企业，每年节省约 200 万元数据成本，并与微众银行合作跑通隐私保护的联邦学习。
The pivot
LLMs arrived and everyone reached for prompts. I saw the gap underneath: a prompt is a wish; nothing makes the wish repeat. So I stopped writing prompts and started building the layer that makes agents reliable.
大模型来了，所有人都在写提示词。我看见的是底下那道缝：提示词是一个愿望，却没有任何东西让愿望重复兑现。于是我不再写提示词，转而去建让智能体可靠的那一层。
AI-Fleet
A personal multi-agent operating system — 400+ skills, context engineering, 38 hooks, routing across models. 46 projects in four months, all standing on one orchestration spine.
一套个人的多智能体操作系统——400+ 技能、上下文工程、38 个钩子、跨模型路由。四个月 46 个项目，全都站在同一根编排脊柱之上。
- Enterprise Data
企业数据 - Kingdee Datawings — platform data operations saving ~2M CNY/year. 金蝶苍穹数据运营，每年节省约 200 万元。
- National Monitoring
国家级监测 - MIIT SME operations monitoring, covering 1M+ enterprises. 工信部中小企业运行监测，覆盖百万级企业。
- Federated Learning
联邦学习 - WeBank collaboration — privacy-preserving multi-party analysis. 与微众银行合作，隐私保护的多方数据分析。
- Industrial IoT
工业互联网 - National dual-cross industrial internet platform data monitoring. 国家双跨工业互联网平台数据监测。
- Finance & Tax AI
智能财税 - Agent systems for automated bookkeeping, tax compliance, invoicing. 面向自动记账、税务合规、发票处理的智能体系统。
46
Projects · 4 months
四个月，46 个项目
四个月，46 个项目
3,300+
Commits shipped
提交的代码
提交的代码
400+
Agent skills authored
编写的智能体技能
编写的智能体技能
2
Patents · 1st / core inventor
专利 · 第一 / 核心发明人
专利 · 第一 / 核心发明人
我的作品 · Works
Infrastructure I build, in production.
我构建的基础设施，都在生产环境里运行。
02 — Enterprise
Cognebula Enterprise
Out-of-the-box enterprise context brain — Graph RAG, a WebGL 3D console, Docker one-click. The company's memory, queryable.
开箱即用的企业上下文大脑——图谱 RAG、WebGL 三维控制台、Docker 一键部署。把公司的记忆，变成可查询的资产。
03 — Vertical Agent
YiClaw 容易记
AI auto-accounting agent for Chinese SMEs — monthly close from 2–4 hours down to 5 minutes. Compliance as a workflow, not a chore.
面向中国中小企业的 AI 自动记账智能体——月度结账从 2–4 小时压缩到 5 分钟。让合规成为一条流水线，而不是一件苦差。
04 — SDK
Armature SDK
Open-source in-process Agent SDK — 12 core contracts, 50 built-in tools, an eval framework. The skeleton other agents are built on.
开源的进程内 Agent SDK——12 个核心契约、50 个内置工具、一套评测框架。别的智能体，都长在这副骨架上。
05 — Marketplace
OpenClaw Foundry
Curated agent-skill marketplace — 37K+ vetted skills with S/A/B/C ratings. Find the right wheel instead of reinventing it.
精选的智能体技能市场——37K+ 经审核的技能，带 S/A/B/C 评级。找到对的轮子，而不是重新造一个。
06 — Orchestration OS
AI-Fleet
The spine under all of it — a personal multi-agent OS with 400+ skills, context engineering, verification gates, and 38 lifecycle hooks.
这一切之下的脊柱——一套个人多智能体操作系统，400+ 技能、上下文工程、验证门、38 个生命周期钩子。
The learning ladder 学习阶梯
四个配套网站——从第一次接触 AI 工具，到跑通你自己的智能体工作流。install.sh. 岗位 Pack——一条 install.sh 装上 PM / 工程师 / 分析师的工作流。
Practice →
我的思考 · Thinking
Five ideas the infrastructure is built on.
支撑这套基建的五个想法。
01
Infrastructure over prompts.
02
The Commander paradigm.
03
Coordination must not exceed the task.
04
Reference over skill.
05
Memory without reflection is a correction log.
「AI 智能体需要的是基建，而不仅仅是提示词。提示词是一个愿望，基建才是让愿望可靠兑现的东西。」
A prompt is a wish; infrastructure is what makes the wish repeat. The orchestration layer — routing, context, multi-model collaboration, verification — is where capability becomes a workflow you can trust twice.
「AI 是战术兵力，人类负责战略决策与结果验收。」
I don't ask the model to decide what matters. I decide what matters, command a fleet of agents to execute, and I own the acceptance gate. The leverage is a clearer command and a harder gate — not a cleverer prompt.
「协调层的复杂度必须与任务的真实复杂度匹配，绝不能超过它。」
Harness depth, swarm breadth, skill count — each must scale with the real task, never past it. A fat harness on a thin task is orchestration debt that quietly degrades quality. Simplicity is a design target, not a leftover.
「只固化公共网络所缺失的部分。」
If a target has a strong public anchor — a canonical algorithm, a standard protocol, a known design language — one sentence of reference beats a skill that re-describes what the model already knows. Build for the constraints the world hasn't written down yet.
「原始存储不是学习。只有经过检索、连接、证据检验并晋升的规则，才算真正学到。」
A system that only appends observations accumulates a correction log, not a mind. Learning is a promotion cycle: retrieve, connect, evidence-gate, and promote only the rules that survive. Everything else stays an observation.