What Is DeepSeek's Super-Agent? How the 'More Autonomous' Coding Agent Actually Works, How Autonomous It Really Is by Its Own Rules, What the Numbers Measure, and How to Run It

What Bloomberg reported, and what DeepSeek actually published

Bloomberg's November 1 story, 'DeepSeek Drops Enhanced Coding Agent With Aim to Rival Anthropic,' is short because the underlying material is a WeChat article and a GitHub repository, not a launch. Its facts, as syndicated here: 'DeepSeek has released a prototype of a more autonomous coding agent as it seeks to challenge Anthropic's popular Claude Code.' The agent 'can proactively monitor group chats among developers for feedback on coding issues,' per 'the DeepSeek Harness team… in a post on WeChat on Saturday.' And the context: 'Earlier this year, DeepSeek created the Harness team to focus on helping build AI agents,' and DeepSeek 'has raised about 100 billion yuan ($15 billion) in its latest fundraising round,' which Bloomberg reported October 22. That is the whole report. 'Rival Anthropic' is Bloomberg's framing; DeepSeek's materials never mention Claude Code.

The primary source is the repository, deepseek-ai/dsh-superagent-prototype, created October 31 at 9:32 UTC, one commit, TypeScript, MIT license, 6,260 stars by the next afternoon. Its README opens with a disclaimer worth reading in full: 'Research prototype for reference only, not an official DeepSeek product. This code grew out of our team's day-to-day development work, with much of the engineering done by coding agents. Treat it as research code, not production-ready software.' It adds that the public version 'is not identical to our internal version or the implementation described in' the WeChat article, 'Exploring Super-agent Design: Towards a More Autonomous Agent System,' and that 'some designs discussed in the article are directions for future development and are not implemented in this prototype.' The WeChat article, as relayed by Chinese outlets (we could not read the original behind WeChat's verification wall), says the same thing from the other side: the release is 'an exploration of architectural ideas for engineering practice, not an introduction to existing product features.'

What Super-agent is, in DeepSeek's own words

From the README: 'dsh-superagent is a proactive coding agent for one GitHub repository, packaged as a bundle for DeepSeek Harness. It listens to GitHub and, optionally, Feishu. When someone allowed asks the agent for work, it opens a work item, works on it in its own agent session and directory, and answers in the original conversation, for example with a pull request. A web panel shows what needs a person.' Feishu is ByteDance's workplace chat, Lark outside China; that is the 'group chats' Bloomberg described, and in the published bundle the Feishu source is present but disabled by default.

The architecture is two layers. A Router (the 'main session') reads every event the repository produces, issues, pull requests, comments, CI results on the default branch, and decides whether each one is a request, a candidate for work, or noise. Workers (one per 'Ticket') do the actual job in their own directory with a full tool set: files, search, shell, background jobs, web, skills, and nested subagents, reaching GitHub through the gh command-line tool under a dedicated agent account. The execution document explains why the split exists: the Router runs with 'a smaller tool set for routing' and five-second shell limits so it stays responsive while workers grind. The WeChat article's headline claim, relayed consistently across Sina, Jiemian, and HK01, is that one person can supervise 'three to five active sessions' this way, because 'the ceiling of a single task depends on model capability; the ceiling of many tasks depends on human attention.'

That sentence is the product thesis. Claude Code, Codex, and Cursor are built around a developer who is present. Super-agent is built around a developer who is absent: the agent is always on, the human is interrupted only when 'a person must act or can use a result,' and the philosophy document judges every feature 'by the attention it costs.' Internally, per the article, DeepSeek runs Super-agent instances continuously on the Harness codebase itself, reproducing and fixing issues, answering code-review comments, and iterating on Harness. One Chinese outlet counted 99 pull requests it believes were merged this way between September 14 and October 29; that is the journalist's inference, not DeepSeek's number.

How autonomous it actually is

Less than the headlines, by design, and that is the most interesting part of the documents. The authorization model is strict about who can start work:

• It acts only when asked, or when a written rule allows it. 'A request counts only when it addresses the agent directly and comes from a person the team allowed: one on an allowed list, or anyone in a conversation the team opened to requests.' Bots can never make a request; a bot on the Source's own list can only produce an 'inferred candidate.'
• Proactive work starts in observe mode. 'Work nobody asked for needs a rule the team wrote, and starts in a mode that only records what the agent would have done.' You switch to auto mode after reading the log of what it would have done and deciding you like it. The documents tell operators to 'start narrow' and widen the allowed list later.
• Prompt injection is addressed structurally. 'The host enforces allowed lists, rules and Mandates. Messages, comments and quoted documents are inputs to a decision; they cannot change those permissions.' A stranger's issue comment cannot promote itself into an instruction.
• It stops by impact, not by effort. 'A local change goes as far as a reviewable result. A change that needs a design decision stops with a clear question.' It is supposed to ask 'one concrete question at a time' rather than bury the question in a report.
• Auto-merge is possible, but fenced. The default instructions allow it only for a pull request the agent opened for its own Ticket, 'when it is ready, the required approvals cover the head commit, and additions plus deletions total fewer than 200 lines.' DeepSeek adds that these 'govern the agent, not GitHub's merge rules,' and tells you to enforce approvals in branch protection.

Then the caveat that matters if you run it. Workers 'run commands without approval as the operator's OS user.' Command wrappers and per-ticket directories 'establish defaults, not a security boundary. A worker can still execute a program by absolute path.' There is no sandbox in the prototype, no per-command approval, and no browser automation or MCP support described anywhere in the architecture document. DeepSeek Harness's own safety file says the runtime 'has not undergone a security audit and must not be treated as secure or production-ready' and recommends 'a disposable virtual machine, container, or dedicated environment.' After last week's trojanized Claude Code extension, that advice applies doubly to an always-on agent with a GitHub token and your shell.

The numbers DeepSeek published, and what they measure

The WeChat article includes a test, and it is narrower than it sounds. As relayed by three outlets that agree on every figure, DeepSeek ran a synthetic coordination benchmark: about 300 messages per run, roughly 20 work requests among them, five runs per setup, timings compressed to a quarter, and, crucially, 'coding replaced by a time-only tool.' Nobody wrote code in this test; it measured how quickly an orchestrator notices, dispatches, and closes requests.

Two more figures: Router decision latency fell from 6.4 seconds at one trigger a minute to 1.6 seconds at forty a minute, because it batches; and delivering follow-ups straight to workers instead of through the Router cut token reads by 37%. DeepSeek labels the results 'preliminary.' The models were not named ('a lighter model' for the Router, 'the strongest model' for Workers). So the honest summary is: Super-agent's coordination layer is about twice as fast as a single long session at managing a queue, at roughly two to four times the token cost, in a test where the work itself was simulated. There is no benchmark of the code it produces, and no independent evaluation of any kind yet.

What model powers it

DeepSeek doesn't say. The prototype resolves an 'Eval' model for the Router and an 'Executor' model for workers from configuration, and the documents name neither. In practice it runs on whatever DeepSeek Harness is pointed at, and Harness's own interface defaults to DeepSeek-V4.1-Flash, the October 17 build, which DeepSeek's model card says is 'optimized for DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy.' On the API that is $0.15 per million input tokens and $0.60 output at off-peak hours, doubling at peak, with cache hits at $0.003, and the test above read 23 million cached tokens per run for a reason. Harness also supports custom model endpoints, so nothing stops you running Super-agent's loop on a different model; DeepSeek's docs even publish an Anthropic-compatible endpoint that maps 'claude-opus' requests to V4 Pro and 'claude-sonnet' to Flash, the reverse direction.

Where it fits: Harness, the Harness team, and the $15 billion

Super-agent is the third act of a six-month story. On May 20, DeepSeek researcher Deli Chen posted that 'DeepSeek is forming a new Harness team to build Code Harness from the ground up' and joked it might be called 'DeepSeek Code'; the job listing asked for experience with Claude Code, Codex, Cursor, OpenCode, Copilot, and Manus, and Chen told Yicai the aim was to 'benchmark against Claude Code.' On August 13, DeepSeek shipped DeepSeek Harness alongside V4 Pro: an MIT-licensed agent runtime with a desktop app for macOS and Windows, a web UI you start with one command, and an 'everything is a plugin' architecture built on a framework called Cordis. It is in 'developer preview' with the README warning, in capitals, that 'there will be compatibility-breaking changes'; GitHub shows 268,000 stars eleven weeks in. September 10 brought V4.1 Flash, October 17 the 1017 build tuned for Harness, and October 22 the reported $15 billion round led by Tencent and CATL ahead of a Shanghai listing.

Read in that order, Super-agent is the Harness team showing what it built Harness for. The agent runtime was the platform; the proactive repository agent is the first thing DeepSeek has published that uses the platform the way the team uses it internally. It is also a statement of where DeepSeek thinks the category goes: not a better chat-in-terminal, but an agent that is a member of the team's GitHub organization.

How to try it

Two layers, both free. DeepSeek Harness itself runs with one command (Node.js 22.19 or newer):

npx @deepseek-ai/dsh web

Or install the desktop app from deepseek.com/harness for Apple-silicon Macs or 64-bit Windows; there is no Linux desktop build. You supply a DeepSeek API key, or point it at another provider. Super-agent then installs as a Harness bundle, and DeepSeek's README is candid that 'setup has many steps, so an agent does it with you': you open your coding agent and tell it to read SETUP.md and set up dsh-superagent with you. You will need a GitHub account for the bot, the gh CLI authenticated as that account, a repository to point it at, and a machine that stays on, since the whole point is that it runs when you don't. Start in observe mode, with a short allowed list, on a repository you can afford to have an unsandboxed agent inside. Use our token counter to see what a 23-million-token day costs at DeepSeek's cache-hit rate versus Claude's; the answer is the reason this design is viable on Flash.

What it isn't

• Not a Claude Code or Codex competitor in form. Those are interactive agents you drive. Super-agent is a background service for a repository. The closest analogues are GitHub Copilot's coding agent you assign issues to, Codex's cloud tasks, and Claude Code's cloud scheduled tasks on GitHub events, none of which watch a team chat.
• Not in the DeepSeek app or API. There is no 'agent mode' to switch on, no model ID, and no hosted version. It is code you run.
• Not a release with a date. DeepSeek has promised nothing beyond the repository. The README says the published code 'is not identical to our internal version.'
• Not evaluated. No SWE-bench, Terminal-Bench, or any coding score exists for Super-agent, and the only numbers are a simulation of coordination. V4.1 Flash, the model it will most often run on, scores 42.3% on Terminal-Bench 4.0 by DeepSeek's own card against Opus 5.5's 66.4% by Anthropic's; our comparison lays the two stacks side by side.

Our read

The design documents are better than most shipped products' documentation, and the thesis, that the limit on agentic coding is human attention rather than model capability, is the one every lab is converging on; Anthropic's Claude Duet and OpenAI's scheduled Codex tasks are the same instinct from the other direction. What DeepSeek has not shown is the part that would make it matter: code quality from an always-on agent on a small, cheap model, measured by anyone but DeepSeek. Until that exists, this is a well-argued architecture and an invitation to run it yourself. We will, on a throwaway repository, and we'll report what it merges. For now Claude Code keeps its place in our rankings, DeepSeek Harness enters the code-editors 'also considered' list as a provisional entry we haven't scored hands-on, and the open-source models ranking keeps describing DeepSeek's weights as 'brilliant open weights you'll rent, not run.'

What's confirmed, what's relayed, and what's unknown

• Confirmed from DeepSeek's repository: everything quoted from the README and the numbered design documents; the MIT license; the October 31 creation date; the dependency on DeepSeek Harness 0.2.1.
• Relayed, consistent across outlets, original unread: the WeChat article's claims and test figures, including 'three to five sessions' and the timing table. We could not open the WeChat original.
• Bloomberg's framing, not DeepSeek's: 'rival Anthropic.' DeepSeek's materials don't name Claude Code; its May hiring post did.
• Unknown: the model DeepSeek uses internally for each role; the Harness team's size; whether any of this reaches the DeepSeek app or API; any date; any coding benchmark; the 99-merged-PR count, which is a reporter's inference.