Post

HKUDS053: OpenOPC - AI-Native Company / Self-Built Self-Run Self-Grown

HKUDS053: OpenOPC - AI-Native Company / Self-Built Self-Run Self-Grown

这是 PENGYI_HKUDS_STUDYMAPHKUDS053

这一次看的是 HKUDS 最新非常值得关注的项目:

1
HKUDS053 -> OpenOPC

我对它的一句话定位是:

1
OpenOPC = AI-native company OS.

不是单个 chatbot。 不是单个 coding agent。 不是单个 workflow runner。

它更像是把多个 AI employee、role、org chart、work item、approval、review、memory、office UI、CLI、channel、skill、tool runtime 组织起来,形成一个“可以自建、自运行、自成长”的 AI company prototype。

这和我们最近一直在讲的 A / B / C 面非常相关:

1
2
3
B 面:coding / research / system design
A 面:真实业务 / 金融组织 / 客户 / 风控 / 交付
C 面:founder / product / team / market / capital

OpenOPC 对我们最重要的启发不是“又一个 agent 项目”,而是:

1
2
真正的 founder / delivery system 不是一个人无限写代码。
它需要组织结构、角色分工、任务流转、复盘学习、审批边界和可视化管理。

Latest Snapshot

本次观察时间:2026-07-14

ItemValue
repoHKUDS/OpenOPC
URLhttps://github.com/HKUDS/OpenOPC
descriptionOpenOPC: Build Your Personal AI-Native Company — Self-Built, Self-Run, Self-Grown
default branchmain
languagePython
stars723
forks110
open issues6
pushed at2026-07-13T15:05:26Z
updated at2026-07-13T16:31:02Z
root structure.opc, config, docs, opc, scripts, skills, tests, pyproject.toml, README.md, README.zh-CN.md

最近十个 commit 很说明问题:

DateCommitMessage
2026-07-134bc18dcfix(ui): stabilize workplace chat scrolling
2026-07-134e7aa75fix(company): stop failed children from wedging the tree and soften the dispatch guard
2026-07-09a040252fix(office-ui): stop progress-row flicker and surface native role replies end to end
2026-07-0935fa717config: give all native org roles the full toolset via empty tools lists
2026-07-09e3bed49perf(office-ui): stop per-token full-tree renders, sleep hidden Phaser loop, cut broadcast churn
2026-07-085aa57e6fix: release comms-blocking parks when replies arrive and give every role final-reply tools
2026-07-084b29b89refactor: unify tool approval into a single engine and cut prompt storms
2026-07-08447516dfix: converge approval park/resume through the work-item state machine
2026-07-07818189afix: guarantee engine replies reach the UI channel and stop same-scope duplicate sends
2026-07-073a2c935fix: deduplicate re-delivered session sends on the client message id

这说明 OpenOPC 已经不是停留在 README 概念层。

它最近的更新集中在:

1
2
3
4
5
6
7
8
UI stability
company runtime failure recovery
office UI rendering performance
role reply routing
tool approval engine
work-item state machine
communication blocking / resume
duplicate message handling

也就是说,它在进入真实 product hardening 阶段。

Core Thesis

OpenOPC README 里给了三个核心词:

1
2
3
Self-Built
Self-Run
Self-Grown

我理解如下。

Self-Built

Self-Built 解决的是:

1
一个复杂目标来了,应该由哪些角色完成?

它不是默认只有一个 agent,而是先搭组织:

1
2
3
4
5
6
goal
  -> org chart
  -> roles
  -> reporting lines
  -> employee assignment
  -> talent reuse or fresh hire

这非常像真实公司。

一个 FICC quant demo 不是只有“写代码”:

1
2
3
4
5
6
7
8
PM / product owner
quant researcher
data engineer
backend engineer
frontend engineer
risk reviewer
documentation lead
deployment owner

如果每个角色都由 agent 或 human-agent hybrid 承担,那么项目就从“个人写脚本”变成“组织化 delivery”。

Self-Run

Self-Run 解决的是:

1
组织搭好之后,任务如何流动?

OpenOPC 的关键不是简单 task list,而是 work-item state machine:

1
2
3
4
5
6
7
8
9
brief
  -> decomposition
  -> work items
  -> dependency DAG
  -> assign owner
  -> execute / delegate / review / integrate / rework
  -> blocker handling
  -> human escalation
  -> final deliverable

这里最重要的是:

1
真实任务无法完全 upfront planning。

所以它需要:

1
2
3
4
5
6
7
dynamic collaboration orchestration
kanban status
role ownership
runnability check
review gate
blocker handling
human approval

这和我们自己的 Transition OS、CV OS、FICC Project OS 完全能接上。

Self-Grown

Self-Grown 解决的是:

1
做完一轮之后,系统怎么变聪明?

它不是把所有成败都归给“整个公司”,而是把反馈归因到具体 role / employee / work item。

1
2
3
4
5
6
execution trace
  -> per-role evaluation
  -> employee experience profile
  -> recurring lesson
  -> shared playbook
  -> next run better

这就是我们一直说的:

1
self-evolution loop

我们自己的项目也应该这样做。

比如:

1
2
3
4
5
FICC rates project v0.7 Cloudflare deploy failed once
  -> root cause: Worker vs Pages config mismatch
  -> lesson: static frontend should use Pages, not Worker wrangler deploy
  -> playbook: Cloudflare Pages setting checklist
  -> next project: directly reuse deploy checklist

这才是真正的 OS 化。

System Design View

OpenOPC 的系统分层可以这样看:

1
2
3
4
5
6
7
8
9
10
11
12
13
User Goal / Brief
  -> Company Builder
  -> Org Chart / Roles / Reporting Lines
  -> Talent / Employee Assignment
  -> Work Item Planner
  -> Work Item State Machine
  -> Role Runtime
  -> Native / External Agent Execution
  -> Tool Approval / Permissions
  -> Comms / Blockers / Human Escalation
  -> Office UI / Kanban / Workspace
  -> Memory / Experience / Playbooks
  -> Deliverable

它的 root structure 也对应这些方向:

Directory / FileMeaning
.opclocal OPC workspace / runtime state / config convention
configsystem, agent, company, channel configuration
docsvisual assets, guides, design docs
opccore Python package
scriptsmaintenance / reset / utility scripts
skillsreusable skill library
teststest suite
pyproject.tomlPython package and dependency definition
README.md / README.zh-CN.mdbilingual project narrative

这就是一个标准的 AI application / agent runtime repo 结构:

1
2
3
4
5
6
7
8
package code
config
docs
skills
tests
scripts
UI assets
quickstart

它不是 toy README。

Office UI

OpenOPC 的 Office UI 很关键。

它不是只给你一个 chat box,而是给你:

1
2
3
4
5
6
7
8
9
10
11
Workspace
Office
Org
Kanban
Execution Progress
Agents tab
Comms tab
Team cockpit
role inspector
talent market
organization editor

这说明它把 agent work 从“文本对话”升级成了“组织管理界面”。

对于复杂工作,chat box 是不够的。

我们做 FICC project、CV project、paper-to-code project、research proposal project,都需要:

1
2
3
4
5
6
7
谁负责什么?
哪些任务卡住了?
哪些需要 review?
哪些已经完成?
哪些要 human approval?
最终 artifact 在哪里?
复盘 lesson 写进哪里?

这就是 Office UI 的价值。

OpenOPC vs AgentSpace vs OpenHarness

之前我们已经看过:

1
2
HKUDS006 -> AgentSpace
HKUDS015 -> OpenHarness

现在看 OpenOPC,可以这样区分:

ProjectLayer
OpenHarnessagent harness runtime: model + tools + skills + permissions + memory
AgentSpacehuman + agents workspace: team collaboration surface
OpenOPCAI-native company: org chart + employees + work items + company mode + self-growth

我的理解:

1
2
3
OpenHarness 更偏 runtime。
AgentSpace 更偏 workspace。
OpenOPC 更偏 organization / company operating system。

三者不是互相替代,而是越来越上层:

1
2
3
runtime
  -> workspace
  -> company

这条线和 HKUDS 近期生态非常一致。

OpenOPC in the HKUDS Ecosystem

现在 HKUDS 的项目已经可以拼成一张更大的图:

1
2
3
4
5
6
7
8
9
LightRAG       -> knowledge memory / graph RAG
RAG-Anything   -> multimodal document RAG
nanobot        -> personal agent shell
CLI-Anything   -> software action / CLI tool layer
OpenHarness    -> agent runtime / harness
AgentSpace     -> human-agent workspace
Vibe-Trading   -> finance / trading research workflow
AI-Trader      -> agent-native trading platform
OpenOPC        -> AI-native company / organization OS

OpenOPC 的位置是最高一层:

1
2
它不是某个 agent 会不会调工具。
它关心的是一个 AI organization 如何自己组队、分工、执行、复盘、成长。

这对我们尤其重要。

因为我们未来不只是要做一个 demo,而是要把多个方向组织起来:

1
2
3
4
5
6
7
8
9
CV delivery
FICC project
AI agent harness
quant research workflow
learning log
interview grilling
public website
GitHub profile
Cloudflare deployment

这些本质上都可以被看成:

1
Pengyi Personal AI-Native Company v0

Connection to Our A / B / C OS

我们之前一直在拆:

1
2
3
A 面:金融业务 / 客户 / 组织 / 资源 / 风控
B 面:coding / research / AI / quant / system design
C 面:founder / product / team / market / capital

OpenOPC 给了一个很直接的工程化表达:

1
2
C 面 founder 不是抽象愿景。
C 面 founder 是把 A 面真实业务和 B 面技术生产力组织成一个可运行公司。

对我们来说:

1
2
3
4
5
6
7
8
A 面提供真实场景:
  bank workflow, credit process, industry chain, FICC, client needs

B 面提供技术能力:
  Python, agent, RAG, quant, frontend, backend, deployment

C 面需要组织化:
  role design, task decomposition, review, human approval, memory, self-growth

OpenOPC 的系统设计正好在讲这个。

How We Can Use This

我们可以把 OpenOPC 的思想迁移到自己的 Research OS:

1. Project as Company

每个项目都可以有“虚拟公司结构”。

比如 Rates Bond Quant

1
2
3
4
5
6
7
Founder / Product Owner
  -> Quant Researcher
  -> Backend Engineer
  -> Frontend Engineer
  -> Documentation Lead
  -> Deployment Engineer
  -> Risk / Public-safe Reviewer

每个角色负责不同 artifact:

1
2
3
4
5
6
7
8
9
quant engine
API design
dashboard UI
README
tests
version log
Cloudflare deploy
CV bullet
interview story

2. CV as Company

CV 也可以被拆成角色:

1
2
3
4
5
6
7
8
Positioning Lead
Evidence Reviewer
Quant Story Owner
AI Project Owner
Banking Experience Owner
English Version Owner
LaTeX Production Owner
Interview Grilling Owner

这样 CV 不是“写一页纸”,而是一个 delivery workflow。

3. Interview as Company

面试准备也能组织化:

1
2
3
4
5
6
WorldQuant interviewer
FICC quant interviewer
AI agent interviewer
banking interviewer
behavioral interviewer
resume truth reviewer

每个角色都拷打不同风险点。

4. Learning as Company

每篇 learning 可以拆成:

1
2
3
4
5
6
source reader
system design analyst
coding analyst
resume storyteller
public-safe reviewer
website publisher

这就是我们现在做 HKUDS learning 的方式。

Why OpenOPC Matters for Founder OS

OpenOPC 最值得我们记住的是:

1
2
founder 不是单点能力。
founder 是组织化复杂任务的能力。

如果我们以后做 C 面:

1
2
3
4
5
6
7
拿资金
带团队
做产品
服务客户
对抗竞品
迭代市场
交付营收

那本质上就是:

1
2
3
Self-Built: 组队
Self-Run: 执行
Self-Grown: 复盘升级

OpenOPC 把这个东西工程化成了一个 AI-native system。

这和我们的长期目标高度共鸣。

Interview Hooks

如果以后面试里讲 OpenOPC,我们可以这样说:

30-second version

1
2
3
I studied OpenOPC as an example of an AI-native organization system.
It goes beyond a single agent by introducing org charts, role assignment, work-item state machines, office UI, review loops, blockers, human escalation, and self-growth memory.
For me, it is useful because it shows how AI agents can move from isolated tools into structured team-based delivery systems.

Chinese version

1
2
3
我研究 OpenOPC 的重点不是“又一个 agent”,而是它把 agent 放进组织结构里:
角色、上下级关系、任务拆解、Kanban、review、blocker、人类审批、经验沉淀。
这对 AI + Quant Research OS 很关键,因为真实研究和真实交付都不是一个 prompt 就结束,而是一个持续协作、复盘、升级的系统。

If asked: why relevant to finance?

1
2
3
Finance workflows are naturally role-based and approval-heavy:
researcher, trader, risk reviewer, compliance, product owner, client-facing team.
OpenOPC's role/task/review/escalation design gives a useful template for building AI-assisted financial research and decision-support systems with human-in-the-loop control.

Active Mock Questions

如果要主动拷打,这篇可以准备这些问题:

  1. OpenOPC 和普通 multi-agent framework 的区别是什么?
  2. 什么是 Self-Built / Self-Run / Self-Grown
  3. 为什么真实任务需要 work-item state machine?
  4. 为什么 Office UI 比 chat UI 更适合复杂 agent work?
  5. AgentSpace、OpenHarness、OpenOPC 的层级区别是什么?
  6. OpenOPC 最近 commit 说明它处于什么阶段?
  7. 它如何处理 blocker、approval、review 和 human escalation?
  8. 这种系统如何迁移到 FICC / Quant Research / Banking workflow?
  9. 如果让你做一个 Pengyi Research OPC,你会设计哪些 role?
  10. 它对 C 面 founder OS 有什么启发?

Final Takeaway

这篇 HKUDS053 的结论很明确:

1
OpenOPC 是 HKUDS 从 agent runtime / workspace 进一步走向 organization OS 的关键项目。

它对我们的价值也很明确:

1
2
我们不能只做散点项目。
我们要把项目、CV、面试、投递、研究、网站、GitHub、部署都变成一个可运行、可复盘、可升级的 Personal AI-Native Company。

这就是:

1
2
3
Pengyi Research OS
  -> Pengyi Delivery OS
  -> Pengyi Founder OS

OpenOPC 是我们理解这个跃迁的一块关键拼图。

Sources

This post is licensed under CC BY 4.0 by the author.