DeepSeek Harness dominates GitHub: the super-agent prototype racks up +16,000 stars in 6 days

DeepSeek Harness dominates GitHub: the super-agent prototype gains +16,000 stars in 6 days

🔎 A GitHub podium monopolized by the "dsh" ecosystem

The GitHub ranking of November 4, 2026 tells a single story: that of DeepSeek. At the top, dsh-superagent-prototype, a repository published on October 31, shows 18,135 stars on the morning of November 4 — ranked #2 worldwide and #2 in TypeScript. Behind it, its foundation, deepseek-harness, peaks at 262,303 stars.

And this is no isolated success. ccch1mneyyy/dsh-Go, the app presented as capable of running the super-agent on a phone, has racked up +8,777 stars in seven days. TokenEconomizer completes the weekly podium with +4,379. For three days now, the "dsh" ecosystem has been taking center stage on GitHub, following the week of Vercel's computer-use and the agents audit, and then the explosion of claude-mem and its persistent memory at 108,000 stars.

Why now? Because this prototype embodies the next stage of coding agents: moving from the reactive agent — you ask, it codes — to the proactive agent, which listens to your tickets and opens pull requests on its own. And the landscape is favorable: according to Upstar, the 443 repositories classified under "coding agents" gained 146,128 stars this week, up 42% from the previous week.

The Essentials

• A vertical curve: dsh-superagent-prototype went from 1,658 stars on October 31 to 18,135 on November 4, 2026 (+16,477), averaging +2,620 stars per day (Trendar).
• A proactive agent: it monitors a GitHub repo (and optionally Feishu), opens a work item for each request, works in an isolated session, and replies in the original thread, typically with a pull request.
• A restraint rule: act only when asked, speak only when a human needs to step in. A deliberate design choice, the opposite of chatty agents.
• An ecosystem, not a project: the prototype is a bundle for DeepSeek Harness (262,303 stars, 32 releases), an "Everything is a Plugin" framework built on Cordis.
• A cautious README: "research prototype for reference only, not an official DeepSeek product". Something to study, not to deploy in production.
• An overheating category: +146,128 stars in one week across 443 coding agent repos (+42% over seven days, Upstar, Oct. 29–Nov. 4, 2026).

Recommended tools

All the projects featured in this ranking are free and open source — which is also what explains how quickly they've been adopted. Here are the six tools to know about this week, and for more, check out our selection of the best AI tools for coding.

The November 4, 2026 rankings in numbers

On November 4, dsh-superagent-prototype is the AI agent repo that gained the most stars over the weekly window, and the second-ranked repo overall on GitHub. Trendar's daily series gives the exact measure of the phenomenon:

Source: Trendar, data as of November 4, 2026.

The counters vary slightly depending on the measurement window: GitHub showed 18,232 stars and 1,987 forks as of November 4, Gitstars recorded +6,001 stars over its 24-hour window, and Git of Anton counted +3,000 stars over the last 24 hours at 12:12 GMT, i.e., +16,259 since it began tracking. Whatever the source: the trajectory is the same.

Over the week of October 29 to November 4, Upstar's "Trending Coding agents" leaderboard sets the tone: dsh-superagent-prototype (+18.2k, entered the ranking at 546 stars), dsh-Go (+8,777), TokenEconomizer (+4,379). The daily Synon leaderboard confirmed the prototype's first place as early as November 2, with +4,939 stars in 24 hours.

My take: the slowdown on November 4 (+1,860) is mechanical and healthy. A repo that passes 18,000 stars in a few days simply has fewer people left to convince. The strong signal isn't the slope, it's holding the #2 spot worldwide three days after the peak — most viral repos fall back within 48 hours; this one is settling in.

dsh-superagent-prototype : the agent that turns your tickets into pull requests

Concretely, dsh-superagent is a proactive coding agent connected to a single GitHub repository. It listens for requests — on GitHub, and optionally on Feishu — opens a work item, works in its own session and its own directory, then replies in the original conversation, typically with a pull request. A web panel shows what's waiting for a human.

The most important detail is that "single repository" part. Where autonomous agents dream of controlling everything, DeepSeek chose a narrow scope: one repository, one agent, work items. The blast radius of an error is contained by design, not by promise.

Second structuring choice: session isolation. The agent never works in your current directory tree but in its own, then proposes the result as a PR. This eliminates the classic failure mode of agents stepping on each other when multiple tasks run in parallel.

DeepSeek had already tested GitHub's appetite with DeepSeek-TUI : le coding agent terminal qui explose sur GitHub avec 5800 stars en un jour. The super-agent takes things to another level: we go from the individual tool to the team workflow, complete with tickets, PRs, and a monitoring panel. That's a different ambition, and apparently, a different resonance.

"Everything is a Plugin": What the DeepSeek Harness Architecture Reveals

The super-agent is not a standalone product: it's a bundle packaged for DeepSeek Harness, DeepSeek AI's open source agent framework. And this framework is massive: 262,303 stars, 31,706 forks, 1,108 watchers, 32 releases, MIT license — all in TypeScript (GitHub, November 4, 2026).

Its architecture boils down to three words: "Everything is a Plugin". Harness is built on Cordis, a modular runtime, and is described in a research paper titled "A Programming Paradigm for Spatiotemporal Composability" (arXiv:2608.25512). In other words: every capability — listening to GitHub, talking to Feishu, coding, opening PRs — is a plugin that composes with the others.

The parent repository's dynamics confirm the platform effect: +73,065 stars in 61 days according to Trendshift (August 23 – October 22, 2026), and another +4,900 stars over the last seven days according to OSS Radar, which ranks it #4 among the 96 "AI agents" repositories tracked this week.

This is the mechanism behind the podium. A core (Harness), a flagship application (the super-agent), and satellites reusing the same foundation — dsh-Go, TokenEconomizer. It's the same scenario as with Prime Agent: GitHub's #1 repo (17,500 stars) that makes coding agents self-improving — and auditable: value is shifting from the agents to the infrastructure that runs them.

A direct consequence of the plugin paradigm: Harness is, in theory, model-agnostic. Nothing rules out imagining bundles built around GPT-5.5 (98.2 on the agentic leaderboard), Gemini 3 Pro Deep Think (95.4), or Claude Opus 4.7 (94.3). DeepSeek is obviously betting on its own family — DeepSeek V4 Pro (Max) stands at 88, the best score among open source LLMs — but the architecture doesn't impose it. To settle the question of which engine to choose, our roundup of the best LLMs for coding details each one's strengths.

If the approach appeals to you, our guide on creating an AI agent starts from exactly these building blocks: a core, tools plugged in as plugins, and a defined scope of action.

The sobriety rule: act only when asked

The most interesting detail in the README isn't technical, it's behavioral: the super-agent only acts when asked to, and only speaks up when a human needs to intervene.

In a market obsessed with maximum autonomy — agents that commit, post, and decide on their own — this is a contrarian bet. DeepSeek is sketching out an agent that is proactive in execution but discreet in initiative: it listens continuously, but only kicks off work when an explicit request arrives in a thread.

The web panel makes this philosophy tangible: it displays "what is waiting for a human." The agent doesn't presume to decide on your behalf; it prepares the work and flags the decision points. This is human-in-the-loop by design, not as damage control after an incident.

I find this choice more mature than the race to total autonomy. The past few weeks have shown us both extremes: self-improving agents that need auditing, computer-use agents that need supervising. In between, an agent that does the requested work and then keeps quiet comes across as plain common sense. The 18,000 stars in just a few days suggest the market approves.

One last signal not to overlook: Feishu. The fact that the prototype can optionally listen to ByteDance's collaborative suite, dominant in the Chinese professional ecosystem, says a lot about the market DeepSeek is targeting first — and about its ambition to align with local usage habits rather than Western standards alone.

dsh-Go and its satellites: the mechanics of an ecosystem

The weekly podium is not three independent successes: it's an ecosystem phenomenon, and it should be read as such.

dsh-Go, second in the weekly ranking with +8,777 stars (Upstar, Oct. 29–Nov. 4), is presented as the app that runs the dsh super-agent on a phone. Synon's daily ranking already had it at 5th place on November 2 (+1,025 stars). The idea — supervising your agents from your pocket — obviously strikes a chord with highly active repository maintainers.

TokenEconomizer, in third place (+4,379), completes the picture. Within just a few days, the "dsh" prefix has become a marker of belonging, almost an internal GitHub keyword: the core attracts extensions, and the extensions feed the core.

A word of caution, however: where a platform explodes, some projects ride the wave. Before adding a star — and even more so before running any code — check the commit history, the license, and the issue activity. A name starting with "dsh" guarantees nothing, neither quality nor any real connection to the DeepSeek ecosystem.

Unslop, Mika, magpie: the rest of the leaderboard isn't sleeping

Outside the "dsh" bubble, the week confirms three underlying trends: AI-assisted design, on-device agents, and multi-model setups.

unslop (heydamir, JavaScript) is the cross-cutting phenomenon of the week: 7,700 stars, including +7.7k in seven days — #1 in the Claude Code category according to Upstar — and four consecutive days trending (bestofgithub.dev). It's a Claude skill that turns a "vibecoded" interface into a genuinely designed one. The name says it all: the market is starting to react to the "slop" aesthetic mass-produced by agents.

Mika (wangzan101, Kotlin, Apache 2.0) embodies the on-device trend: an AI agent that runs directly on Android, with no cloud. 3,718 stars as of November 2 (+1,121 in 24 hours, Synon), around 6,200 by November 4 (Upstar, #2 among AI agents for the week). The promise: local latency and data that never leaves the phone.

magpie (yetone, Go) is already a heavyweight on the leaderboard with 29,300 stars and +610 per day, backed by the v0.1.2888 release. Its promise fits in a single sentence: "Every agent's model. One place." Codex on DeepSeek, Claude Code on Kimi, straight from the menu bar. We're entering the era of on-demand model routing — to pick the right engine for the task, see our comparison of the best LLMs for coding.

Finally, Pika-H3-Flex (Pika-Labs, Python) marks the arrival of open-source media inference in the top trending: video and audio generation, 1,400 stars, +750 in 24 hours. Also worth noting on the daily leaderboard for November 2: patch-gg/pixelfx (+1,185) and haydenschmitty/floorplan (+824), a codebase mapping tool.

We track these moves every week in our AI news section.

Should you install dsh-superagent-prototype today?

Not in production, yes in a sandbox. The README says so itself: "research prototype for reference only, not an official DeepSeek product". The code comes from the team's day-to-day development, largely written by coding agents, and matches neither DeepSeek's internal version nor the implementation described in the research paper. Last commit: October 31, the day of publication. It's a snapshot, not a maintained product.

Three precautions before testing:

A dedicated environment. The agent listens to GitHub and, optionally, Feishu — so it needs a machine that runs continuously. A small VPS is more than enough for the listening and the web panel; at Hostinger, a few euros a month will do (November 2026, check hostinger.com).

A throwaway repository. Never hook the prototype up to your main repository. Create a test repository, with branch protection and systematic PR review. The agent opens pull requests — let it do so, but review everything.

Reading the code. That's the whole point of an MIT-licensed TypeScript prototype: the code is readable, and the README openly admits it's largely agent-generated. Reading it is already a lesson in how DeepSeek structures its agents in practice.

For day-to-day work, the safe bets remain the proven reactive agents — Claude Code, Codex, or OpenCode: 8 million devs, 172K GitHub stars — the open source coding agent overtaking Claude Code and Codex. The super-agent is not a replacement: it's a preview of what the "ticket → PR" workflow will become once it matures.

My advice: install it to understand the paradigm, not to save time. The value of this repository, today, is educational.

❌ Common mistakes

Mistake 1: confusing a research prototype with an official product

The name "deepseek-ai" and the 18,000 stars give the impression of a finished product. That's false: the README states in black and white that it is not an official product, that the code differs from the internal version, and that it should be treated as research. The solution: follow deepseek-harness releases and wait for stabilization before any serious use.

Mistake 2: hooking a proactive agent up to your production repository

An agent that opens PRs continuously can turn your Git history into a field of ruins if a bug slips into its loop. The solution: a dedicated sandbox repository, branch protections, and one simple rule — no agent PR gets merged without human review. The "waiting on a human" panel exists precisely for that.

Mistake 3: confusing stars with maturity

+42% growth in the "coding agents" category in one week also means that attention shifts faster than quality. 18,000 stars in a few days measure hype, not a user base or reliability. Before adopting a trending project: check the license, commit frequency, quality of the issues, and run an isolated test.

❓ Frequently Asked Questions

Is dsh-superagent-prototype an official DeepSeek product?

No. The README states it explicitly: "Research prototype for reference only, not an official DeepSeek product". The code comes from the team's day-to-day development, largely produced by coding agents, and does not correspond to DeepSeek's internal version. It's a research artifact under the MIT license, published on October 31, 2026.

How many stars has the repository gained exactly?

According to Trendar, it goes from 1,658 stars on October 31 to 18,135 on November 4, 2026, i.e., roughly +16,500 stars, with a peak of +5,870 on November 2. GitHub shows 18,232 stars and 1,987 forks as of November 4. The discrepancies between sources correspond to different measurement windows.

What's the difference with Claude Code or OpenCode?

Claude Code, Codex, or OpenCode are reactive agents: you start a session, you ask, they code. dsh-superagent is proactive: it continuously listens to a GitHub repository (and Feishu), turns a request into a work item, and replies in the thread with a PR. The two approaches are complementary, not competing.

What model does the super-agent run on?

The README doesn't mandate it: the prototype is a bundle for DeepSeek Harness, whose "everything is a plugin" architecture is model-agnostic. Logically, DeepSeek relies on its own family — DeepSeek V4 Pro (Max), the best open-source LLM currently available with a score of 88 — but the plugin paradigm theoretically allows other engines.

What is Feishu, mentioned in the README?

Feishu is ByteDance's collaborative suite, known internationally as Lark and dominant in the Chinese professional ecosystem. The super-agent can listen for requests on both GitHub and Feishu — a clear signal about DeepSeek's priority market and its intent to align with local practices.

✅ Conclusion

In less than a week, the DeepSeek Harness ecosystem has just proven that a solid framework and a well-thought-out prototype can capture all of GitHub's attention: 18,000 stars for the super-agent, 262,000 for the foundation, and a podium monopolized for three days straight. What comes next will play out in upcoming releases — in the meantime, keep the sandbox habit, and follow our coverage of AI news as well as our selection of the best AI tools so you don't miss anything.