---
id: "lib-2026-11-05-lobehub-hands-its-r-and-d-to-agents"
type: "library"
title: "LobeHub hands its R&D to agents"
title_original: "一家8万Star的AI公司，决定让AI接管自己"
lang: "ZH"
org: "白鲸实验室 (on 36Kr)"
source_type: "Practitioner post"
date: "2026-11-05"
week: "2026-W45"
url: "https://nosov.ai/library/lib-2026-11-05-lobehub-hands-its-r-and-d-to-agents/"
original_url: "https://www.36kr.com/p/4056040123945860"
tier: "B"
tier_name: "Useful, limited"
tier_reason: "Baijing Lab feature republished by 36Kr on the open-source agent platform LobeHub, combining company-described practice with the reporters' own timed tests; useful first-hand detail, but small tests and vendor framing."
summary_basis: "full-text"
original_minutes: "10"
tags: ["ai technology", "enablement", "coding agents", "agent orchestration"]
---

# LobeHub hands its R&D to agents

Original title: 一家8万Star的AI公司，决定让AI接管自己

Tier B (Useful, limited): Baijing Lab feature republished by 36Kr on the open-source agent platform LobeHub, combining company-described practice with the reporters' own timed tests; useful first-hand detail, but small tests and vendor framing.

Summary based on the full text.

## In short

A feature by Baijing Lab (WeChat), republished on 36Kr on 5 November 2026, on LobeHub, a Chinese-founded open-source agent collaboration platform with about 84,000 GitHub stars that runs its own R&D through its agents. A supervisor agent assigns work to executor agents on Claude Code and Codex; one recent PR went from creation to merge with a single human action, the merge click. Humans mostly read multimodal acceptance reports instead of code. The reporters reproduced three workflows and found them working, with manual gaps in the feedback loop.

## Key numbers

- ~12 s: time from an untagged chat message to a new task appearing in the reporters' first test
- 2 min 46 s: time for Claude Code to build a small web page with dark mode, styles and tests
- 2 min 31 s: time for Codex to re-run 17 checks and review that page, finding no reproducible problem
- 3 months: time the LobeHub team says it has not started a local development server
- 0.65: responsibility-sensitivity setting of the test agent that decides whether a chat message is its job

## What it reports

- PR #21852, submitted by the account lobehub-coder on 29 October, was reviewed by a Codex bot; the two exchanged several rounds in 18 minutes at 4 a.m. on 31 October, and the coder declined one finding with reasons. A human engineer merged it on 1 November.
- A supervisor agent ("Nasa") reads the task board, investigates issues and assigns them to executor agents. It was introduced after agents posting directly to the group chat produced "all noise".
- Each task has an executor slot and an owner slot; the owner may be a person or another agent.
- The team says it has not started a local dev server for three months; the Acceptance tool (open-sourced in September) turns requirements into observable checks with screenshots, recordings and logs, and a person passes or returns items one by one.
- Reporter test 1: an untagged message ("nobody has sorted this week's topic list") created a task in about 12 s, the supervisor accepted in under 40 s, the executor ran 26 steps and delivered in under 4 minutes; a joke about a bookshelf was correctly declined.
- Reporter test 2: Claude Code built a page in 2 min 46 s; after the desktop window was closed, Codex took over on the same task card and reviewed it in 2 min 31 s (17 checks); 5 files, 515 lines.
- Reporter test 3: Acceptance produced an 8-item report in about 14 minutes; reporters returned one item and Claude Code changed one CSS line. Returning feedback still needed manual copy-paste and a CLI login, and 3 items showed as failed after an extra empty run.
- Context: DeepSeek's "Super-agent" article, which says one person can supervise about 3 to 5 agents.

## How they got it

Interviews and GitHub records for the company's own practice; three small hands-on tests by the reporters in a Feishu group, with timings and screenshots.

## Limits

- Company practice is described by the company; no data on defect rates, rework or cost.
- Each reporter test was run once on a toy task.
- The reporters note that checking the page directly took about 19 s, faster than reading the report (about 63 s).

## Why it matters for companies

This is a concrete picture of a supervisor-and-worker agent setup in daily use: one router agent, a written role document, two seats per task, and human sign-off on acceptance evidence rather than on code. The weak points the reporters found (feedback loops that still need manual steps, permission set-up, and no evidence on long-term memory) are where similar internal pilots will also stall. Security, architecture and final quality stay with people.

## Key passages in translation

> 单任务的上限取决于模型能力，多任务的上限取决于人的注意力。

The ceiling on a single task depends on model capability; the ceiling on many tasks depends on human attention. (DeepSeek, quoted in the article)

> 他们甚至称，团队已经几乎不看代码了，核心看的就是 Agent 的多模态验收报告。

They even say the team hardly looks at code any more; what it mainly reads is the agents' multimodal acceptance reports.

> 报告里的"打回重跑"按钮只会复制一段指令，还得手动粘贴、发送

The report's "return and rerun" button only copies an instruction, which still has to be pasted and sent by hand.

Original: https://www.36kr.com/p/4056040123945860