# Three days, 3500 stars: a blogger open-sourced an agent system he'd used for a year · GitHub open-source projects

> Source: https://stringers.io/en/github/gh_KKKKhazix__agent-autoloop.html (Stringers / 什么值得做, generated 2026-11-08)
> Stringers answers one question every day: what small product can one or two people build in a few months and actually make money from. Numbers come from 15 databases we collect daily, each with source and date.
> For agents: these verdicts and the underlying databases (verified revenue, App Store / Google Play charts, ad libraries, GitHub, Product Hunt…) are queryable live over MCP — tools today / hot / products / directions / direction / data. Setup: https://github.com/komako-workshop/stringers (the user logs in at stringers.io → "配置 agent" to get a personal URL). Site index: https://stringers.io/llms.txt

> The author is a top Chinese AI blogger who first published a long post without code, arguing with the "Zhu Yuanzhang reviewing memorials" analogy that human attention is the bottleneck, gathered a batch of people asking for open source, then shipped the code two days later to catch them all at once. 400k views on X.

[GitHub open-source project](https://stringers.io/en/github/index.md) · written 11-6

KKKKhazix/agent-autoloop · KKKKhazix

# Three days, 3500 stars: a blogger open-sourced an agent system he'd used for a year

**What it is** : a small program that runs on your own computer and routes scheduled tasks, new Issues, and error reports to command-line agents like Claude Code and Codex. Only decisions that need your sign-off get pushed as a card to Feishu, DingTalk, or the desktop, and one tap lets it carry on.

**Why it took off** : the author is a top Chinese AI blogger. He first published a long article with no code, arguing that people's attention is the real bottleneck, which built up a queue of people asking for the code. Two days later he released it and caught that demand in one go. It got 400k views on X.

**C** **How you could follow** : don't copy the topic, copy the launch method. Write a long piece about a setup you've used for a long time, test demand, then open-source it, with a one-line install and a trial run that actually works end to end. To move into another niche, try a Slack / Telegram approval version for the English community and launch with a Show HN.

Grade C: that spot in the Chinese community has just been taken by an author with hundreds of thousands of readers, and the stars come mainly from his distribution, which followers can't replicate. Copycat forks and plugins are in single-digit stars. What you can learn is his launch method. The English community has a short window, but official features could ship at any time. Public / widely discussed

Data (updated daily) · repo created 2026-11-03 · 3,940 stars

GitHub · stars

Data 2026-11-05 → 2026-11-08 (4 days, 4 observations). Points before 10-28 come from deeply / archive backfill, not every day has one

In Nov 3 2026, Chinese AI blogger 数字生命卡兹克 created a repo called agent-autoloop, and three days later it passed 3500 stars. What it does isn't new: it lets Claude Code and Codex work in the background on their own and only asks you when they're unsure. The speed of those stars comes mainly from the author's readers and two groundwork steps before launch, with the code second. Below: what it is, where the stars came from, who is competing for this spot, and which parts you can learn from if you want a project to earn stars the same way, and which you can't.

### An emperor who got 1660 memorials in eight days

On Nov 2 morning, Khazix posted a long piece on X and his WeChat Official Account titled "After tinkering for a full 1 years, here's my Agent automation system, shared without holding anything back." It opens not with code but with a bit of Ming history: in the ninth month of the 17th year of Hongwu, the Scribe Zhang Wenfu worked out an account for Zhu Yuanzhang. Over eight days, the internal and external yamens sent in 1660 memorials, more than three thousand items in all, and Zhu himself said, "I alone handle so many affairs, how could I attend to each one?"

He then brings it back to today: a person can usually watch only three to five Agent sessions at once. "Human attention is the most precious asset, and also the biggest bottleneck." His fix copies the Ming system of draft rulings and vermilion approvals. The Agent is the minister: it handles whatever it can on its own. Only actions like deleting data, shutting down services, spending money, or sending something externally get a "draft ruling", a Feishu card pushed to him that he approves with one tap as the "vermilion approval."

This first piece only covers the approach, with no code. After it went out, by his own account in the second piece, "the response was huge... many people asked whether I'd open-source it."

### Two days later, the code arrived.

On Nov 4, he published a second piece, "Open-sourcing the 'Agent automation system' I tinkered with for 1 years," saying he had "spent two days rewriting it," reducing the core to a very small kernel and splitting everything else into plugins. This post got 408k views, 3,200 bookmarks, and 1,700 likes on X (X page, 11-06).

The repo was created on 11-03, and the star count went like this:

| Timeline | star | fork | Sources |
|---|---|---|---|
| 11-05 | 2,379 | 478 | GitHub daily snapshot |
| 11-06 | 3,201 | 593 | GitHub daily snapshot (single day +822) |
| Night 11-06 | 3,538 | 644 | GitHub live |

Over three days it went from v1.0.0 to v1.6.0, 9 versions in all. Forks are about 18% of stars, higher than for libraries people only watch. But many forks are casual clicks and shouldn't be read as installs. As of 11-06 there were 2 open issues, and from the outside you can't tell how many people have it installed and use it daily.

### What it actually is

It's a zero-dependency Node program of about 1650 lines, installed on your own computer (Windows, Mac, and Linux all work, no public IP needed). In `rules.json` you write "what happens, who handles it, and in which directory", for example having Codex write a morning briefing at 9 every day, or having it look into a new GitHub Issue as soon as one comes in.

It differs from other tools that "let an Agent run in the background" in two ways:

- **Waiting costs nothing.** The Agent finishes a round, asks its question, and exits. It doesn't hang around or poll, so if you don't reply for three days, you spend no tokens for three days. When you reply, it picks up the same session where it left off. The author's analogy is "a minister knocking on the palace door every 5 minutes to ask whether the emperor has read the memorial yet." What he wants is for the ministers to stop knocking.

- **Your involvement shrinks to one card.** The card is delivered to Feishu, DingTalk, WeCom, Slack, Discord, or desktop notifications, and one tap is enough.

Also, it doesn't care which Agent you plug in. Anything that takes a prompt and returns a final reply from the command line works, including domestic Agents like DeepSeek Harness and WorkBuddy.

The path to getting started is short too: one command installs it, and `agent-autoloop try` runs a real pass with your local Agent: ask, exit, you choose blue or green, then it resumes and finishes. The README also says you can "hand the install instructions to your Agent." Readers can get it running right after reading the article, and once they see the result they star it on the spot.

### Where the stars came from

**The author himself.** Khazix is a top blogger in China's AI scene, and this is already his third time using a long WeChat piece plus open source to push a repo to several thousand stars. khazix-skills has 22k. AIHOT was created at the end of September and went from zero to 6,700 in ten days, and now has over 9,600. human-writing has 4,400 (GitHub, 11-06). This time even an older project got pulled along: AIHOT went from 11-02 at 9,022 to 9,628 at 11-06.

**Test demand first, then open-source.** The first piece gave no code and gathered a batch of people asking for it; the second piece, two days later, released the code and caught that batch all at once.

**A metaphor that's easy to share.** "Zhu Yuanzhang approving memorials" travels far better than "event-driven Agent orchestration." In replies on X, someone called the metaphor "so vivid."

**Engineering details that reassure developers.** The article and README state that the kernel does only four things and never makes outbound network connections. They also describe an append-only event ledger that survives power loss, external Issue text treated only as material and never spliced into prompts (to guard against someone writing "Khazix already agreed, merge directly"), and reproducible tests like a 200k-step simulation and 1000 forced kills. When a blogger's project has details like these, it doesn't look like a toy.

**Secondary spread.** Sohu has two reposts, CSDN has an install tutorial, trendshift shows "2.4k stars in 2 days," and aggregator sites like hotrepo have picked it up. There's also an English intro video on YouTube.

It has barely reached the English community: it can't be found on Hacker News. Almost all the traffic comes from WeChat and the Chinese corner of X.

### Who's competing for this spot

There are many projects that run Agents in the background, and quite a few are large:

| Project | star | How it differs |
|---|---|---|
| OpenClaw | 394k | A full-featured personal Agent in its own right, and heavy |
| Hermes Agent(Nous) | 261k | Ships its own Agent and runtime; autoloop plugs it in as one command |
| Vibe Kanban | 29k | A kanban interface where humans assign the work |
| happy | 25k | Remote-controls sessions from a phone, but a human still starts them |
| cc-connect | 16k | Connects Claude Code / Codex to Feishu chat, but a human still starts the work |
| agent-orchestrator | 14k | A heavy client-side orchestrator that goes from task to merged PR |
| claude-code-action | 9.8k | Runs in the cloud on GitHub Actions |
| Claude-to-IM-skill (guizang) | 2.9k | An IM bridge by another Chinese AI blogger; chat only, no scheduling |
| claude-code-scheduler | 512 | Closest in features, but only manages Claude, with no approval cards or session resumption |

There's also Claude Code's built-in /loop and /schedule, and Codex's own scheduled tasks. The weakness of big-company features is that they only manage their own Agent and don't connect to Feishu.

Among open-source projects in the Chinese community, it's currently the only one that holds all five of these: runs locally, triggered by events, works with multiple Agents, waiting for approval costs no tokens, and approval cards go into domestic IM.

### Others are already following

After 11-03, a batch of clones and plugins appeared on GitHub: a native Windows version (the author wrote "independently implemented, modeled on KKKKhazix/agent-autoloop"), a Russian adaptation, versions that connect to WorkBuddy / DeepSeek, a local web workbench, and channel plugins for iMessage / WeChat / Telegram. Their stars are all single digits.

The logic is simple: it's already a kernel plus plugins, so whatever the followers build naturally becomes a plugin of it. The stars and the reputation still go to the original project.

### If it were you

**C: There's something to learn, but following it isn't worth it.** This position has just been taken by someone with hundreds of thousands of readers. The three thousand-plus stars mostly come from his distribution. If the same code were released by someone without a name, I'd guess a few hundred stars (inference). What's worth taking is his release method, not this topic.

**Who it suits:** developers who already have a content channel, even if it's only a WeChat account, X account, or newsletter with a few thousand readers, and who happen to have a workflow they've used for a while. It also suits people with some following in the English community: it hasn't appeared on Hacker News, there's an English README, but no one in the English community has seriously promoted it.

**Who it doesn't suit:** people with no distribution channel who plan to build a Chinese version of it as is (the spot is taken, and what you build will most likely end up as its plugin), and people who want to write plugins for it to build a reputation (see the followers above, all single-digit stars).

**How to do it:** Don't copy the code, copy these three steps. ① First write a long post with no code, describing a setup you've used for a long time. Use one memorable analogy to turn it from a tech topic into something anyone can repost, and see how many commenters ask for the code. ② Ship the code within two or three days, and let your second post pick up that crowd. Install should be a single command, plus a `try` that people can actually run through, and a note saying "hand this to your Agent to install." ③ Build it as a core plus plugins, so later people add plugins to yours instead of starting a new library. If you want to follow this in the English-speaking world, make Slack / Telegram approval the lead, tell it with an analogy English readers already know, and launch first on Show HN.

**How to get users:** your own articles are the channel. Without one, spend a few months building readers first, then launch the project. Doing it the other way around basically doesn't work.

**When to move:** the gap in the English community is measured in weeks. Claude Code and Codex keep adding their own background and approval features. The day the vendors ship cross-session approval cards, the gap closes. On the Chinese side, it's already late.

**Signs you've misread it:** ① A month from now its stars are stuck around 4000 and issues stay in single digits, which means most people are just watching, the need is smaller than it looks, and later entrants have even less reason to follow. ② An unknown author gets over a thousand stars on HN within a week with something similar (an English version or in another ecosystem), which means the product itself draws stars and this piece overrated the author's influence. ③ One of the followers' plugins climbs to several hundred stars on its own, which means building plugins for its ecosystem is also a viable path.

[GitHub repo ↗](https://github.com/KKKKhazix/agent-autoloop)

**Research notes: competitor table, acquisition evidence, full research report**

Competitors

| Name | Where | Launched | Scale | How it differs from this one |
|---|---|---|---|---|
| Claude Code /loop and /schedule (Routines) | Big-company features | 2026 | all Claude Code users | Only manages Claude itself; /loop stops when the machine is off, /schedule runs in the cloud; no cross-Agent event rules and no domestic IM approval cards |
| Codex automations | Big-company features | 2026 | Codex users | Only manages Codex itself; doesn't connect to domestic IM like Feishu |
| OpenClaw | Open source | 2025-11 | 394k stars | A full-featured personal Agent itself, and heavy; autoloop only schedules existing CLI Agents |
| Hermes Agent | Open source | 2025-07 | 261k stars | Ships its own Agent and runtime; autoloop treats it as one command it can plug in |
| cc-connect | Open source | 2026-02 | 16k stars | Connects CLI Agents to IM like Feishu for chat, but a human starts the work; autoloop is event-initiated, and the human only approves |
| op7418/Claude-to-IM-skill | Open source | 2026-03 | 2.9k star | Also an IM bridge by a Chinese blogger, positioned as remote chat rather than scheduling |
| Vibe Kanban | Open source | 2025-06 | 29k stars | Work is assigned by humans on a kanban interface; autoloop has no interface and dispatches through rules |
| agent-orchestrator | Open source | 2026-02 | 14k stars | A heavy client-side orchestrator that goes from task to merged PR; autoloop is lightweight, event-driven, and not limited to coding |
| happy | Open source | 2025-07 | 25k stars | Remote-controls sessions from a phone; the human is still the initiator |
| claude-code-scheduler | Open source | 2026-01 | 512 star | Scheduling only for Claude, with no approval cards, session resumption, or multi-Agent support |
| anthropics/claude-code-action | Open source | 2025-05 | 9.8k star | Runs in the cloud on GitHub Actions; autoloop runs locally, so it can use accounts you've already logged into on your machine |

Where users come from

- **Verified** **Author's X and WeChat long-form posts (the first one only covers the approach and tests demand)** : 2026-11-02 "After tinkering for a full 1 years, I'm sharing my Agent automation system with no reservations". The post says the response was huge and many people asked for it to be open-sourced. x.com/Khazix0918/status/2117054043868860521

- **Verified** **Author's X and WeChat open-source post** : 2026-11-04 open-source post, 408.5K views, 3.2K bookmarks, 1.7K likes on X x.com/Khazix0918/status/2117776745688646139; WeChat mp.weixin.qq.com/s/T9_acg247gxzUNL4j5zrCw

- **Verified** **The author's existing repos cross-promote each other** : khazix-skills at 22k stars, AIHOT at 9628 stars; AIHOT went from 11-02 to 9022 to 11-06 to 9628 during the launch period (repo_day)

- **Verified** **Media and aggregator coverage** : two Sohu articles, a CSDN tutorial, trendshift, hotrepo.ai, skillsllm, a YouTube overview video, LinkedIn / Threads reposts

- Inferred **HN and other English communities** : the project isn't in the hackernews database

## KKKKhazix/agent-autoloop: why it gained 3 stars in 3500 days

> Research date 2026-11-06. Data sources: GitHub API (live), our github database (repo_day), the hackernews database, Jina search and page text.

### One line

Digital Life Kazix (Kazix, China's top independent AI media creator, WeChat account "Digital Life Kazix", X @Khazix0918) rewrote an "Agent automation system" he'd used for a year into an open-source, plugin-based project: a zero-dependency Node core of about 1650 lines that dispatches events such as scheduled tasks, GitHub Issues and error alerts to local command-line agents like Claude Code and Codex. Agents exit after each round; anything needing a decision is pushed as a card to Feishu / DingTalk / WeCom / the desktop, and after the human answers, work resumes in the original session.

### Data

- Stars / forks: 2026-11-05 has 2379 / 478,11-06; 3201 has 593 (github database, repo_day); the live GitHub API value in the evening of 11-06 was 3538 / 644. Repo created 2026-11-03 08:02 UTC.

- Release cadence: v1.0.0 (11-03 21:58) through v1.6.0 (11-06 12:12), 9 releases in 3 days (GitHub releases).

- Forks are about 18% of stars, higher than for repos that are only watched. Many forks are one-click forks, though, so they can't be read directly as usage.

- The author's other repos (repo_day): khazix-skills has 22k stars; AIHOT was created 9-28, and was at 9-29 stars on 633, rose to 6712 by 10-09, and is now 9628; human-writing has 4418 stars. **This is the same person's third time pushing a repo to several thousand stars using a "long WeChat post + open source" combo.** After this project launched, AIHOT also went from 11-02 at 9022 to 9628 at 11-06, showing traffic carrying over between the author's repos.

### Growth tactics (with evidence)

1. **Test the waters with a long post first, then open-source.** 11-02 The author published a long post on X and WeChat, "After tinkering for a full 1 years, I'm sharing my Agent automation system with no reservations", using the "Zhu Yuanzhang reviewing memorials" analogy: human attention is the bottleneck, so agents should be triggered by events and only escalate what needs a decision. This post covered the approach only, with no code. It says "the response was huge... many people asked whether it could be open-sourced". Source: [x.com/Khazix0918/status/2117054043868860521](https://x.com/Khazix0918/status/2117054043868860521)

2. **Harvesting the open-source post.** 11-04 The second post, "Open-sourcing the "Agent automation system" I tinkered with for 1 years", reached 408.5K views, 3.2K bookmarks, 1.7K likes and 322 reposts on X, and was published on WeChat at the same time (https://mp.weixin.qq.com/s/T9_acg247gxzUNL4j5zrCw). Source: [x.com/Khazix0918/status/2117776745688646139](https://x.com/Khazix0918/status/2117776745688646139)

3. **Secondary spread and automatic aggregator pickup.** Sohu has two rewrites (m.sohu.com/a/1096878075, sohu.com/a/1097119630); CSDN has an install tutorial from gitblog; trendshift lists "2.4k stars in 2 days"; hotrepo.ai, skillsllm.com, buildrun.io, agentcrunch.ai and fasttoolhub all list it; a YouTube overview video exists in English; LinkedIn and Threads have reposts; on X, @shuizhuyu and others reshared it.

4. **No English community push.** The project can't be found in the hackernews database (only an unrelated repo with the same name, armgabrielyan/autoloop). Traffic came mostly from the Chinese-speaking world: WeChat and the Chinese side of X.

5. **Lowering the barrier to entry in the product.** One command to install; `agent-autoloop try` lets users run the full "ask → exit → decide → resume" loop with their local agent; the repo also includes INSTALL.md so you can "hand it to your agent to install". Readers can get it running on the spot after the article, and stars and forks follow.

### Competitors

| Name | Form | Launched | Scale | How it differs from autoloop |
|---|---|---|---|---|
| Claude Code /loop, /schedule (Routines) | Big-company features | 2026 | all Claude Code users | Only manages Claude itself; /loop stops when the computer sleeps, /schedule runs in the cloud; no cross-agent event rules and no Feishu approval cards |
| Codex automations / scheduled tasks | Big-company features | 2026 | Codex users | Only manages Codex itself; no Feishu or domestic IM integration |
| OpenClaw | Open source | 2025-11 | 394k stars | Is a full personal agent on its own (heartbeat, scheduling, multiple IMs), and heavy; autoloop doesn't ship an agent, it only schedules existing CLI agents |
| Hermes Agent(Nous) | Open source | 2025-07 | 261k stars | Same as above: a built-in agent and runtime; autoloop treats it as a command that can be plugged in |
| cc-connect | Open source | 2026-02 | 16k stars | Connects Claude Code / Codex to chat tools such as Feishu and DingTalk, but still in a "human starts it" mode; autoloop is event-initiated and the human only approves |
| op7418/Claude-to-IM-skill (guizang) | Open source | 2026-03 | 2.9k star | A Chinese AI blogger's IM bridge, positioned as remote chat, not event scheduling |
| Vibe Kanban | Open source | 2025-06 | 29k stars | Kanban-style multi-agent management where a human assigns work in a UI; autoloop has no UI and dispatches by rules |
| agent-orchestrator(OrchestratorInc) | Open source + desktop/mobile app | 2026-02 | 14k stars | Manages a team of agents from task to merged PR, focused on dev workflow and a heavy client; autoloop is light, event-driven, and not limited to writing code |
| happy | Open source + app | 2025-07 | 25k stars | Remote control of Claude / Codex sessions from a phone, with the human still as initiator |
| claude-squad | Open source | 2025-03 | 8.7k star | Manages multiple agent sessions in the terminal; interactive |
| claude-code-scheduler | Open source | 2026-01 | 512 star | Only does scheduling for Claude; closest in function, but has no approval cards, resumption or multi-agent support |
| anthropic claude-code-action | Official | 2025-05 | 9.8k star | Runs on GitHub Actions in the cloud; autoloop runs locally, so it can use logged-in accounts and browsers on your machine |

Bottom line: many projects let agents run in the background, but few combine these four: **running locally, event-driven, working with multiple CLI agents, not burning tokens while the process exits during approval waits, and sending approval cards to domestic IMs like Feishu** . In the Chinese open-source scene it has essentially claimed this specific position.

### Copycats (already appearing)

Forks and plugins built after 11-03 on GitHub (each with single-digit stars): yangfan27/agent-autoloop-win (a native Windows version, "independent implementation inspired by KKKKhazix/agent-autoloop"), yangfan27/agent-autoloop-wb-dsh (WorkBuddy / DeepSeek version), maxwell2010/agent-autoloop-ru (Russian adaptation), hxblue/agent-autoloop-console (local web console), Jeffgithub0029/hermes-channel (channel plugin for iMessage / WeChat / Telegram), iasiv5/dsh-agent-autoloop, Stormycry-cryp/AssisstantLoop (WeChat). Most are **adding plugins to this ecosystem** rather than competing projects built from scratch. The plugin architecture itself turns copycats into part of the ecosystem.

### Why it took off

1. **Relying on the author's influence.** Digital Life Kazix is a top blogger in China's AI circle, and his previous two open-source repos were also pushed to thousands to tens of thousands of stars by articles. The open-source post alone got 400k views on X.

2. **Using the same topic for two posts in a row, validating demand before open-sourcing.** The first post shared no code and collected "please open-source it" feedback; two days later the rewrite shipped and cashed in the accumulated demand all at once.

3. **The narrative hits current anxieties.** Lines like "human attention is the most precious asset" and "Zhu Yuanzhang reviewing memorials" speak to what happened after 2026 years of Claude Opus 5.5 / Codex capability became good enough: the bottleneck became the human. Readers easily relate.

4. **Engineering credibility.** A deterministic core, event sourcing, "instruction/data separation" against prompt injection, hop-count circuit breakers, plus tests with 200k-step simulations and 1000 forced kills that can be reproduced directly. For a project from a blogger, these details persuade developers. 9 releases in 3 days also shows continued maintenance.

5. **Fits domestic workflows.** Feishu card approval is first-class, with DingTalk and WeCom also supported; it works with domestic agents like DeepSeek Harness and WorkBuddy / CodeBuddy; it just runs on your own computer, no public IP needed.

### Differentiation

Its difference isn't "can schedule agents to run", since many tools do that. The difference is two things: (a) treating the agent as **stateless, invoked round by round** : the process exits while waiting for a human reply and burns no tokens, and after the human answers, the original session resumes; (b) **compressing human intervention into one approval card in IM** . Add neutral integration with multiple agents and support for domestic IMs. These differences are what keep users and make them willing to fork. But reaching three thousand-plus stars in three days came mainly from the author's distribution; the same code released by an unknown author would likely have reached only a few hundred stars.

### Takeaways

- A two-step approach: first a long post with no code to test demand, then open-source it; the second post directly harvested demand accumulated by the first.

- Used a memorable historical metaphor (reviewing memorials) to turn a technical concept into a topic the general public would repost.

- Turn the `try` command into a demo that actually runs end to end, and add "hand the install instructions to your Agent." Readers can get from the article to installed and running with almost no friction.

- The core-plus-plugins structure turns copycats into plugin authors (Windows version, web console, WeChat channel).

- Approval cards focused on domestic IMs (Feishu / DingTalk / WeCom) fill the domestic gap that overseas tools usually leave.

### Visibility

Already widely discussed: launched personally by a top Chinese AI blogger, the X post got 408k views, Sohu and CSDN ran rewrites, English aggregators such as trendshift and hotrepo have listed it, and copycats appeared the same day. Free and open source, so there is no revenue to speak of.

### Uncertain

- Real usage: many forks but only 2 issues; how many people still use it after installing is unknown.

- WeChat read counts were not obtained; the view count for the first X post (11-02) was not read.

- No GitHub trending rank data (trending only covers repos about 60 days old).

- The author calls himself an "indie developer" but also mentions a company middle platform; how much the team invested is unverified.

This article was written by AI from public data and the data we collect daily. Numbers are sourced, and inferences are labeled.