DeepSeek published the code for a coding agent that doesn’t wait to be asked. On October 31, 2026, the Chinese AI lab’s Harness team posted dsh-superagent-prototype on GitHub under the MIT license, alongside a design article titled “Exploring Super-agent Design: Towards a More Autonomous Agent System.” Within a day, the repository had passed 4,000 stars, a sign of how hungry developers are for open coding-agent infrastructure they can run themselves.
Super-Agent is built for a team, not a single developer. It watches one GitHub repository and, optionally, Feishu group chats, ByteDance’s workplace messaging app. When an authorized person asks for something, or when a rule the team wrote allows it, the agent opens a work item, does the job in its own isolated session and directory, and replies in the original conversation, often with a pull request.
That shape is the whole point. Most coding tools today wait for a prompt. You open a terminal, type a request, and watch the agent work. Super-Agent flips the arrangement: it sits on the repository like a junior teammate in the channel, scanning for work it has been cleared to take. DeepSeek states the design goal plainly in its user manual: “Take as little human attention as possible, and give as much value as possible.”
How the DeepSeek Super-Agent open source agent works
Under the hood, the system splits into two layers. An always-on coordinator, called the Router, listens for requests and keeps shared context across the project. It hands each approved task to a Worker agent, which does the actual coding in a separate session and then reports back. A web panel shows what needs a human’s eyes.
Setup is not trivial, and DeepSeek doesn’t pretend otherwise. Running it takes a dedicated GitHub account for the agent, Node.js 22.19 or newer, pnpm, Git, tmux, Python 3.8 or newer, and GitHub CLI 2.48 or newer, plus a model provider configured in DeepSeek Harness, the company’s open agent framework. Only repositories on github.com are supported; GitHub Enterprise Server is out of scope for now.
DeepSeek’s own numbers come with a large asterisk. In a synthetic test described in the design article, roughly 300 messages containing about 20 work requests, the Super-Agent took a median of 12.4 minutes per task from request to result, with the slowest 5 percent finishing in 15.7 minutes. Those are DeepSeek’s own preliminary measurements, not independent benchmarks, and the repository publishes no comparison against Claude Code or OpenAI’s Codex.
A research prototype, not a product
Here’s where the story gets more careful. DeepSeek labels the code a “research prototype for reference only, not an official DeepSeek product,” and warns that it is not production-ready. Much of the engineering was done by coding agents themselves, the README says, and the code “may still contain bugs.”
That warning matters more than usual right now. The prototype’s workers run commands without approval as the operator’s system user, and the code depends on a preview version of DeepSeek Harness. Security researchers who have examined the docs flag the missing sandbox as the biggest risk: an agent that can act on its own, with real system access, is exactly the kind of tool that needs tight guardrails before it touches a real codebase.
No price. No product page. No benchmark.
There is also a gap between what the code does and what the design article describes. DeepSeek says the published prototype differs from the version its own team runs internally and from some of the designs in the article, which remain future directions rather than shipped features. Anyone downloading it this week is getting a blueprint to study and test, not the system DeepSeek uses day to day.
Why this matters
Bloomberg framed the release as a move against Anthropic’s Claude Code, one of the most widely used AI coding tools, and the timing fits. Anthropic is heading toward a public listing, Claude Code is a core part of its revenue story, and an open, free alternative, even a rough one, changes the conversation about who controls the agent layer. GitHub, meanwhile, showed its own multi-repository agent app at its Universe conference last week, so the race toward agents that work across teams, not just terminals, is clearly underway.
Still, the honest read is narrower than the headlines. A study cited by The Next Web, called CooperBench, found coding agents averaged about 30 percent lower success rates when working as a team than when each worked alone. That study didn’t test DeepSeek’s agent, but it points at the real bottleneck: an agent that watches a whole team can just as easily drain everyone’s attention as save it, if its pull requests need heavy correction.
Context cuts the other way, too. DeepSeek built the Harness team earlier in 2026 to compete with coding services from US labs, and Bloomberg reported it raised about 100 billion yuan, roughly $15 billion, ahead of a planned public listing. Yet last month, US civilian federal agencies were ordered to remove DeepSeek from government systems, so the same company courting open-source developers is locked out of parts of the US public sector.
What to watch is concrete. Does the repository keep growing, or does interest fade once developers hit the rough edges? Does anyone ship a hardened fork with a real sandbox? And does Anthropic answer with pricing or feature moves of its own? The code is free, the idea is public, and the attention budget it promises to save is now the thing everyone will measure.
Frequently asked questions
What is DeepSeek Super-Agent?
Super-Agent is an open-source, proactive coding agent published by DeepSeek’s Harness team on October 31, 2026. It watches one GitHub repository, takes requests from authorized team members, and answers with work such as pull requests.
Is DeepSeek Super-Agent free to use?
Yes, the code is published under the MIT license, so anyone can download, run, modify, and redistribute it. DeepSeek stresses it is research code, not a finished product, and it comes with no hosted service or support.
How is Super-Agent different from Claude Code?
Claude Code is an interactive tool a developer drives from a terminal, while Super-Agent is an always-on service that sits on a repository and picks up work on its own. Bloomberg framed the release as a challenge to Anthropic’s Claude Code.
Is DeepSeek Super-Agent safe to run in production?
No. DeepSeek says the prototype is not production-ready, may contain bugs, and differs from the version its own team runs internally. Its workers run commands without approval, so it should only be tested in isolated environments.
Sources: Bloomberg, HelpNetSecurity, The Next Web, and DeepSeek’s GitHub repository for dsh-superagent-prototype.