Skip to content
View xhqing's full-sized avatar
🎯
Focusing
🎯
Focusing

Block or report xhqing

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
xhqing/README.md

🌐 English | 简体中文

Hi there 👋 I'm Huaqing Xu

Building an AI Agent organizational structure & Open-source AI Agent team capabilities

Agent Development · Quant Trading · Agentic Trading · Social Media · Advertising

🇭🇰 Hong Kong  ·  🐍 Pythonista  ·  ❤️ Open Source & Geek mindset

followers   Sponsor  


🚀 Flagship Project

An OPC (one-person company): I personally run the whole agent organization — every member except me is an AI Agent.

The full roster (CEO → four direct reports + three squads · 15 specialized agents) — every member with their project repo, persona name, host machine, team membership, and subprojects:

Title · Agent project Persona Host machine pi-team/pi-alt-team Subprojects (🌐 public · 🔒 private)
dev-agent
Full-stack Engineer
🗺️ Atlas macmini&macbook pi-team 🌐 ghostty-launcher — one-click Ghostty terminal summon from the VSCode status bar (zero-dependency extension, macOS)
🌐 codef — CLI that opens VSCode full-screen with the target window on top (bash + osascript, macOS only)
🌐 pi-harness — independent fork of the Pi agent harness: coding-agent CLI, agent runtime, unified multi-provider LLM API, TUI components (TypeScript monorepo)
🌐 ghostty — independent fork of the Ghostty terminal: v1.3.1 baseline + clipboard-image-paste patch, cloud-built via GitHub Actions
🌐 cmux — independent fork of the cmux terminal (Ghostty-based macOS terminal for AI coding agents): upstream release tag as the baseline plus our own close-confirmation behaviour, kept in sync with upstream tags
🌐 turn-diff — per-turn git diff for pi: records a baseline each turn, reports what that turn changed (`/turn-diff` in a session, `turn-diff` in a terminal), pure git
🌐 mp4-player — VSCode video player: plays videos with audio in an editor tab; bundled ffmpeg WebAssembly for decoding & conversion
🌐 android-cmux — the cmux client for Android: watch terminal and coding-agent sessions from a phone; talks to cmux over its wire protocol only (Kotlin + Compose)
🌐 XPilot — Xray-core proxy-node manager CLI (Python)
quant-agent
Quant Strategist
📐 Mark macmini&macbook pi-team&pi-alt-team 🔒 NewsScoreStudy
assistant-agent
Executive Assistant
👩‍💼 Kit macmini&macbook pi-team 🌐 xhqing — this profile repository
🌐 vscode-copy-anywhere — macOS background helper: files copied in the VSCode explorer can be pasted into Finder / WeChat / any app (converts VSCode's private clipboard format to the system one)
🌐 ark-key-pool — pi extension that keeps a pool of Volcano Ark (Agent Plan) API keys and switches to the next one automatically when quota or validity runs out (cooldown shared across processes)
🌐 macmini-tmux — one command to reach tmux on the Mac mini: session names are numbered automatically, reconnect by name, or attach to the most recent session
🌐 agents-infra — the open-source distribution of the infrastructure every agent runs on: its `pi/agent/` directory mirrors the global `~/.pi/agent/` (rules, skills, extensions, scripts, docs) byte for byte, published one-way from the authoring MacBook (name history: CapabilityManagerAgent → TeamToolkit on 2026-11-05 → agents-infra on 2026-11-11); its `apps/` directory mirrors the configs of the terminal apps the agents themselves run in (ghostty, cmux) — the live files under `~/.config` are authoritative, and one publish script refreshes the mirror
social-agent
Social Media Manager
🥂 Gatsby macmini&macbook pi-team 🌐 blog — personal blog "藏经阁", a docsify site published at xhqing.github.io/blog; articles sync automatically from the social-agent article folder, drafts included
trading-agent
Day Trader
📈 Victor macmini&macbook pi-team 🌐 TradingProofs — machine-checked Lean 4 + Mathlib proofs for fixed-fraction trading math (compounding growth, strong law, Kelly optimal f*)
BackendEngineerAgent
Backend Engineer
🔨 Anvil macmini&macbook pi-team 🌐 CC-BRIDGE — Claude Code upstream bridge framework (Node.js)
NeuralCoreAgent
AI Algorithm Engineer
🧠 Ada macmini&macbook pi-team 🌐 AgentCortex — deep-reasoning engine rules (Infinite / Rapid / Incisive)
🌐 ToolCallingRules — LLM tool-calling ruleset (22 rules / 7 categories) injected into system prompts to cut function-calling argument errors and unsafe retries
apply-agent
Apply Optimizer
🎯 Hopkins macmini&macbook pi-team —
LegalAgent
Legal Agent
⚖️ Justin macmini&macbook pi-team —
ProductStrategistAgent
Product Strategist
🧭 Scout macmini&macbook pi-team —
ProductProducerAgent
Producer
🛠️ Wright macmini&macbook pi-team —
SiteBuilderAgent
Site Builder
🏗️ Mason macmini&macbook pi-team —
GrowthMarketerAgent
Growth Marketer
📣 Buzz macmini&macbook pi-team —
DigiVendAgent
Sales Ops
💰 Vendy macmini&macbook pi-team —
DataAnalystAgent
Analyst
📊 Echo macmini&macbook pi-team —

The team is organized into three squads by domain, with four direct reports outside them all:

  • 👩‍💼 Kit (Executive Assistant) — direct report, not in any squad. The owner's first assistant and the team's generalist: picks up and handles almost anything — not a deep specialist in a single domain, but a multi-skilled all-rounder who manages whatever comes along.
  • 🥂 Gatsby (Social Media Manager) — direct report, not in any squad. Runs the owner's own WeChat community "AI前沿跨界交流群" (an AI-frontier, cross-industry exchange group): ideas, rules, topic calendars and retrospectives come from the agent; the owner guides and executes. Private community only — public-channel marketing stays with Buzz.
  • 🎯 Hopkins (Apply Optimizer) — direct report, not in any squad. Owns the full work-intake loop: gig & job hunting, application engineering (proposals & résumés), funnel tracking (application → reply → talk / interview → win / offer), pricing & salary tests — making the applications convert.
  • ⚖️ Justin (Legal Agent) — direct report, not in any squad, serving all squads cross-team. Owns everything contracts-and-payments across the whole team: contract drafting, staged-payment design, due diligence, evidence chains, dispute response.
  • 🛍️ Digital-product sales squad — the six-stage pipeline (Scout → Wright → Mason → Buzz → Vendy → Echo) turns one-time work into recurring passive income — see the dedicated section below.
  • 📈 Investing & trading squad — Mark develops and backtests quant strategies; Victor is the day trader.
  • 🧰 Infrastructure squad — Anvil handles backend development; Atlas handles full-stack projects (including the proxy toolchain) plus the testing and regression safety nets of its own subprojects; Ada covers AI Agent algorithms and reasoning engines.

🛍️ Digital-product sales squad: an artifact-relay pipeline

Running an end-to-end digital-product business: research → produce → build → traffic → sell → analyze. 🔁 A growing roster of single-purpose AI Agents; six of them form this artifact-relay pipeline, turning one-time work into recurring passive income.

How the pipeline hands off — six AI Agents pass artifacts down the line:

  1. 🧭 Scout sizes up trends, market, rivals and profit potential → an Opportunity Report (what to sell, to whom, at what price) → Wright.
  2. 🛠️ Wright turns the report into a finished product — Prompt packs, templates, ebooks, assets → Mason.
  3. 🏗️ Mason builds the storefront — independent site, landing pages, payment integration — and hands off ready-to-sell buy links → Buzz.
  4. 📣 Buzz repackages it into channel-specific traffic content with buy links (X · IG · YouTube · RedNote · Zhihu · Bilibili, etc.) → Vendy.
  5. 💰 Vendy lists, prices, fulfills and supports across storefronts, closing the sale → sales data → Echo.
  6. 📊 Echo attributes results to decisions and folds lessons into a playbook → feeds back to whichever pipeline stage needs it · 🔁 closing the loop.
Pipeline: ① Scout → ② Wright → ③ Mason → ④ Buzz → ⑤ Vendy → ⑥ Echo

⚙️ How the org runs

The interesting part is not the roster — it's the machinery between the agents:

  • 📦 Artifact handoff. Work moves between agents as defined deliverables, not chat: Scout's Opportunity Report feeds Wright's product build, Mason's storefront (site, landing page, or payment setup) hands Buzz ready-to-sell links, Vendy's sales data lands in Echo's attribution engine. Each agent owns an output contract; the pipeline is the sum of those contracts.
  • 🧰 Shared capability layer. The capabilities every agent runs on — global skills and rules — live in a single authoritative source, with an open-source mirror as the distribution channel. This used to be an agent's job (Prometheus); since 2026-11-05 that role is retired and the project is a plain project under Kit — agents-infra — whose mirror is generated and published one-way by a script. A fix lands once and ripples to all 16 agents.
  • 🧠 Reasoning as infrastructure. Specialized reasoning-engine rules (Infinite / Rapid / Incisive, by Ada) plus an evaluation system sit under every agent's thinking — reasoning depth is treated as shared infrastructure, not per-agent improvisation.
  • 🐞 Quality gates with test/implementation separation. Test-case definition sits with Atlas (dev-agent): cases are written on the subproject worktree before implementation starts, and Workers treat them as read-only; CI still holds the final say — implementers can't bend test cases to fit their code, and "fix A, break B" gets stopped before merge, not after. This strict separation applies only to Atlas's own subprojects that involve complex software work; other projects are out of scope.
  • ⚖️ Contracts & payments as a first-class function. Justin structures staged payments, due diligence and evidence chains for anything the org sells or signs — the org treats legal/payment design as an engineering discipline, not an afterthought.

By the numbers: 15 specialized agents · 3 business squads · 4 direct reports · 6 pipeline stages · 200+ public repositories.


🛠️ Tech & Focus


📬 Connect

GitHub   WeChat ID   Email   RedNote ID


💖 Sponsor

If my open-source work helps you, consider sponsoring — this keeps me motivated to keep creating more and better AI Agents and other Projects.

Sponsor

⭐ Star the repos you like · 🔱 Watch the team grow

Popular repositories Loading

  1. Informer Informer Public

    Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting (AAAI'21 Best Paper)

    Python 31 3

  2. trading-agent trading-agent Public

    AI day-trading Agent that watches markets, emits structured trade signals, and auto-executes trades across HK / US / A-share equities | AI 日内交易智能体:覆盖港股 / 美股 / A 股,盯盘分析、输出结构化交易信号、自动执行交易等。

    Python 6 1

  3. gridtrader gridtrader Public

    Grid-trading strategy development & backtesting toolset (Python / backtrader): A-share daily-K via baostock, CSV-driven grid backtests, Pipenv-managed. | 网格交易策略开发及回测工具:baostock 拉 A 股日 K、CSV 驱动网格回测、…

    Python 5 1

  4. HKIPO HKIPO Public

    港股打新科学策略 | Data-driven HK IPO subscription strategy — 7-dimension scoring, veto rules, and multi-account tactics optimized for the post-2025 HKEX new regulations era.

    3

  5. CC-Bridge CC-Bridge Public

    Claude Code upstream bridge framework (GLM/DeepSeek/MiMo/Kimi/Qwen) — routes Claude Code to third-party models via spoofed-whitelist IDs, per-model thinking levels configured in the bridge config, …

    JavaScript 3

  6. social-agent social-agent Public

    Social media manager agent — Gatsby runs the owner's social matrix (YouTube / Xiaohongshu / WeChat / X / GitHub): content, cadence, and retrospectives | 社媒运营智能体:Gatsby(社媒运营官)负责社媒账号矩阵的内容、节奏与运营复盘;私域核…

    Python 3