Agent infrastructure engineerAI Agent 基础设施工程师

Agent memory is a solved problem. I solved it. 经济系毕业,供应链出身,现在在造 AI 的基础设施

Tong Wu · 吴桐. I build agent infrastructure — memory systems, multi-agent orchestration, MCP servers. 304 stars, 6,444 installs/month across six npm packages, written and maintained solo. I came from supply chain; turns out agents have the same coordination problems. 吴桐 · Tong Wu. 我构建 Agent 基础设施 —— 记忆系统、多 Agent 编排、MCP 服务器。六个 npm 包,304 Star、月均 6,444 次安装,每行独立编写。从供应链转型而来——Agent 的协调失败和物流的是同一类问题

304
GitHub stars · AgentRecallGitHub Star · AgentRecall
6
npm packages, shipped solo独立发布的 npm 包
6.4k
downloads / month月下载量
57
controlled experiments (Plywood)对照实验(Plywood)
Available now现在可入职 Düsseldorf · EU Blue Card · open to relocate杜塞尔多夫 · 欧盟蓝卡 · 可 relocate EN (C2) · DE (C1) · 中文 · 日本語
What I'm shipping正在交付

Production agent infrastructure — designed, written and maintained by one person.生产级的智能体基础设施 —— 由一个人设计、编写、维护。

Not prototypes. Live npm packages with real users, versioned releases, test suites and benchmarks. Each one is the tool I needed and couldn't buy.不是原型。是有真实用户、版本化发布、测试套件与基准的线上 npm 包。每一个,都是我当时需要却买不到的工具。

Flagship旗舰v3.4.27

AgentRecall

agent-recall-mcp · -sdk · -cli · -core

A learning loop for AI agents, not just a memory store. Every correction — "no, not that version" — becomes a weighted record, recalled later, auto-promoted to a cross-project insight. Five cognitive memory layers, FSRS-lite decay + RRF fusion, local markdown, zero cloud.给 AI agent 的学习闭环,而非记忆存储。每次纠正 —— "不是那个版本" —— 都成为带权记录,后续召回,自动晋升为跨项目洞察。五层认知记忆、FSRS-lite 衰减 + RRF 融合、本地 markdown、零云。

304★ GitHub 6.4k installs/mo 31 tests
TypeScriptMCPMonorepoVector retrieval
Sole author独立作者
Web data网络数据v0.8.4

novada-mcp

one server · all web data

Search, scrape, crawl, proxy and AI research in a single npx install. 13 composable tools where competitors ship one capability each — or 69 tools that bloat the context window. One API key, structured output.搜索、抓取、爬取、代理与 AI 研究,一条 npx 命令搞定。13 个可组合工具,而竞品要么只做其中一件、要么塞 69 个工具撑爆上下文。一个 API key,结构化输出。

1.2k installs/mo 40+ tools工具 195+ countries国家
TypeScriptMCPWeb automation
Sole author独立作者
Infra基础设施v1.8.3

Proxy4Agent

bestproxy4agents · residential proxy MCP

Routes an agent's HTTP traffic through 2M+ residential devices to bypass Cloudflare / Akamai / DataDome. Multi-provider adapter layer (Novada, BrightData, Smartproxy, Oxylabs), sticky sessions, city-level geo-targeting. CI + Smithery listed.把 agent 的 HTTP 流量经 200 万+ 真实住宅设备转发,绕过 Cloudflare / Akamai / DataDome。多供应商适配层(Novada、BrightData、Smartproxy、Oxylabs),粘性会话,城市级地理定位。已接入 CI 与 Smithery。

4 provider adapters供应商适配 CI Smithery-listedSmithery 收录
Node.jsMCPNetworking
Sole author独立作者
Orchestration编排框架v0.11.0

AAM — Agent Autonomous Mode

multi-agent orchestration with budget control带预算控制的多 Agent 自主编排

FSM state management, hard budget caps, Telegram alerts, Codex sandboxing. Opus orchestrator → parallel Sonnet workers → independent reviewer → verifier. State in files, survives context compression. First live run: 33 dispatches, 80+ artifacts.FSM 状态管理、预算硬上限、Telegram 告警、Codex 沙箱。Opus 编排 → Sonnet Worker 并行 → 独立 Reviewer → Verifier。状态存文件,穿越 context 压缩。首次实战:33 次派发、80+ 产出物

TypeScriptFSMMCPCLISDK
Sole author独立作者
Research · internal研究 · 内部工具v0.5.0

Plywood Protocol

NL → pseudocode selective compiler自然语言 → 伪代码选择性编译器

Compiles agent briefs to pseudocode so agents follow if/else and for...of more reliably than numbered steps. 57 controlled experiments, effect size d=3.62 (length-controlled). Systematic tasks improve 6–11 points; analytical tasks stay prose. Not open source — internal team protocol, deployed via MCP server + Claude Code hook.将 Agent 指令编译为伪代码,使 Agent 遵循 if/elsefor...of 的可靠性远超编号步骤。57 组对照实验,效应量 d=3.62(长度控制后)。系统性任务提升 6–11 分;分析性任务保持散文。非开源——团队内部协议,通过 MCP Server + Claude Code Hook 部署。

PythonMCPRCT designAblation
Sole author · 57 experiments · production-deployed独立作者 · 57 组实验 · 已部署生产

I also build full-stack web apps and automation (this site, internal tooling, Tampermonkey suites) — useful, but not the headline.我也做全栈网页应用与自动化(这个网站、内部工具、Tampermonkey 套件)—— 有用,但不是重点。

The path here一路走来

A trajectory, not a job hop.一条轨迹,不是跳槽履历。

5+ years across 4 countries: operations at scale → automation → agent infrastructure. The roles look unrelated; the throughline is the speed at which I absorb a new system and ship in it. Blue dots = engineering.5 年以上、跨 4 国:大规模运营 → 自动化 → 智能体基础设施。岗位看似无关,串起它们的是「进入新系统、立刻交付」的速度。蓝点代表工程。

2026 · NOW
AI Automation & Infrastructure Engineer code
Novada Labs · Remote / Düsseldorf, DE

Own the company's agent-facing infrastructure end-to-end — from the MCP product layer to the internal automation that runs the team.端到端负责公司面向 Agent 的基础设施 —— 从对外的 MCP 产品层,到驱动团队运转的内部自动化。

  • Agent infrastructure: architect, ship & maintain the MCP product stack — 6 npm packages (304★, 6.4k installs/mo), 40+ tools, multi-language SDKs (TS/Python/Go).Agent 基础设施:架构、发布并维护 MCP 产品栈 —— 6 个 npm 包(304★、月均 6.4k 安装)、40+ 工具、多语言 SDK(TS/Python/Go)。
  • Automation pipelines: built the daily competitive-intel pipeline (auto-reports to GitHub), n8n workflow nodes, and benchmark harnesses — a 200-URL, 6-dimension eval framework that exposed a competitor's 33% anti-bot failure rate.自动化流水线:搭建每日竞品情报流水线(自动报告到 GitHub)、n8n 工作流节点、基准测试框架 —— 200-URL、6 维度评测框架,发现竞品 33% 反爬失败率。
  • Multi-agent orchestration: designed AAM — a 4-tier FSM stack (Opus orchestrator → parallel Sonnet workers → independent reviewer → verifier) with budget caps and overnight autonomous runs.多 Agent 编排:设计 AAM —— 四层 FSM 架构(Opus 编排 → Sonnet Worker 并行 → 独立 Reviewer → Verifier),带预算上限与隔夜自主运行。
  • Production infra & integrations: built the EU AI Gateway (Claude API relay, GDPR-compliant, AWS Lightsail), shipped the Claude Plugin Marketplace listing, and completed LangChain / CrewAI integrations.生产基础设施与集成:构建 EU AI Gateway(Claude API 中转、GDPR 合规、AWS Lightsail)、发布 Claude Plugin Marketplace 列表、完成 LangChain / CrewAI 集成。
Sep 2025 — Jan 2026 · 5 mos
Station Process Excellence Manager opsautomation
Amazon · Nutzen, DE · On-site
  • Built a Tampermonkey automation suite that cut shift-prep time by 40% — where the engineering instinct first met operations at scale.自研 Tampermonkey 自动化套件,班前准备时间减少 40% —— 工程直觉第一次撞上大规模运营。
  • Owned last-mile delivery & station ops at 100% safety/quality compliance; drove Six Sigma / Kaizen across stations.主导末端配送与站点运营,100% 安全/质量合规;跨站推行六西格玛 / Kaizen。
Jun 2024 — Aug 2025 · 1 yr 3 mos
Supply Chain Manager — Delivery ops
Xiaomi Deutschland · Germany · On-site
  • End-to-end fulfillment & logistics for B2B/B2C omnichannel key accounts (DE & NL) — ≈350k units/year, in 3 languages.B2B/B2C 全渠道核心客户的端到端履约与物流(德/荷)—— 年发货 ≈35 万件,三语协作。
Dec 2023 — Jun 2024 · 7 mos
EU Certification Specialist compliance
Adecco · Düsseldorf, DE · On-site
  • Owned regulatory workstreams across consumer-electronics lines; coordinated labs, notified bodies & product teams in 3 languages.主导消费电子产品线的法规合规工作流;三语协调测试机构、公告机构与产品团队。
Oct 2022 — Feb 2024 · selective role
Tutor & Research Assistant Python · Z-tree
University of Freiburg · Freiburg, DE
  • Selected to tutor 3 required courses — Advanced Macroeconomics I & II, Economic Policy, Finance. Led tutorials, graded, mentored.被选拔担任 3 门必修课助教 —— 高级宏观 I/II、经济政策、金融。带辅导课、阅卷、辅导学弟妹。
  • Research assistant in applied econometrics & AI-labor research (Z-tree, data analysis).应用计量经济学与 AI 劳动市场研究助理(Z-tree、数据分析)。
2020 — 2023 · Education
M.Sc. Economics & Finance (VWL)
University of Freiburg · Grade 3.49 / 4
  • Speaker at the 2023 Faculty of Economics graduation — Outstanding Student Representative.2023 经济学院毕业典礼发言人 —— 优秀学生代表。
  • Thesis: UBI in the AI Era — socio-economic impacts of the GPT ecosystem. Econometrics, applied finance & ML.硕士论文:《AI 时代的全民基本收入 —— GPT 生态的社会经济影响》。计量、应用金融与机器学习。
2014 — 2018 · Education
B.A. Business
Jishou University · Jishou, CN
  • Vice President, 13th Student Union · Head of Discipline Inspection, 12th Student Union.第十三届学生会副主席 · 第十二届学生会纪检部部长。

Papers, certs & references →论文 · 证书 · 推荐信 →

What I do well我擅长的

Engineering at the core, operations as the moat.工程是内核,运营是护城河。

The combination is rare: someone who can write the agent stack and has run physical operations at Amazon scale. I know what the automation is actually for.这个组合很少见:既能亲手写出整套智能体栈,又真的在亚马逊量级跑过实体运营。我清楚自动化到底是为了什么。

01

AI & Agent EngineeringAI 与智能体工程

The core. Proven by 6 shipped packages.内核。由 6 个已发布的包背书。
TypeScriptNode.jsMCP serversSDK / CLI designMulti-agent orchestrationRAG / Vector DBFSM agentsPrompt engineeringLangChainPython
02

Strategic Operations & Supply Chain战略运营与供应链

The moat. Amazon & Xiaomi, at real scale.护城河。亚马逊与小米,真实量级。
Last-mile logisticsS&OP / demand planningDSP managementSix Sigma Green BeltOmnichannel opsSAP / ERP
03

Data Intelligence & Analytics数据智能与分析

The connective tissue.连接组织的部分。
SQL / PandasEconometricsTime-series forecastingA/B testingPower BI / TableauR
Off-screen另一面

What I do when I'm not shipping.不写代码的时候,我在做什么。

I read the I Ching the way some people read systems-design docs — both train you to see structure under noise.我读《易经》,方式像别人读系统设计文档 —— 都是在训练「看穿表象、识别结构」。

I Ching · BaZi · Eastern systems thinking易经 · 八字 · 东方系统思维

A 3,000-year-old binary state machine for decision-making under uncertainty — surprisingly close to how a well-designed FSM agent reasons about the world.三千年前的二进制状态机,用于不确定性下的决策 —— 和一个设计良好的 FSM 智能体推理世界的方式惊人地接近。

Strategy策略Cycles循环State machines状态机Pattern recognition模式识别
䷀ ䷁ ䷂ ䷃ ䷄
64 hexagrams · 4096 transitions64 卦 · 4096 种转换
Let's talk聊聊

Let's build something real.一起做点真东西

Open to AI infrastructure, agent platform, and AGI research engineering roles. Available now · Düsseldorf-based · EU Blue Card · open to relocate.关注 AI 基础设施 / 智能体平台 / AGI 研究工程类岗位。现在可入职 · 常驻杜塞 · 持欧盟蓝卡 · 可 relocate。