By Thorsten Meyer | November 1, 2026 | ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
On Saturday evening, DeepSeek’s Harness team posted a technical article on WeChat and put a piece of code on GitHub. By Sunday morning, the headlines read: DeepSeek launches an autonomous coding agent to challenge Anthropic’s Claude Code. Bloomberg, the South China Morning Post and Investing.com all framed it as a race with the big US coding tools.
Then you open the GitHub page. The first line says: “Research prototype for reference only, not an official DeepSeek product.”
Both things are true, in their way. DeepSeek did publish a working design for a coding agent that does not wait for your prompt. It did not launch a product you can buy, and it did not publish a single benchmark against Claude Code or OpenAI’s Codex. Here is what is actually in the box, what the idea gets right, where the fine print matters, and what it means for companies and developers.
Key takeaways
- What was published: on 31 October 2026, DeepSeek’s Harness team released a design article (“Super-agent Design Exploration: Toward a More Autonomous Agent System”) and the open-source repository dsh-superagent-prototype on GitHub (MIT license, TypeScript, version 0.1.0, one commit).
- What it is not: the repository calls itself a research prototype, “not an official DeepSeek product”. There is no price, no product page, no API announcement and no benchmark against Claude Code or Codex. The rivalry is the headlines’ framing.
- The idea: DeepSeek’s team says one person can usually keep track of only three to five agent sessions at once. Its answer is an always-on agent that listens to GitHub and, optionally, Feishu team chats, and hands each job to a separate worker that answers with a pull request.
- The leash: the released code acts only for people on an allowed list or under rules the team writes. Proactive work starts in an “observe” mode that only records what the agent would have done. It does not merge its own pull requests.
- The fine print: the workers run with full access as the operator’s own user account and can read the home folder. DeepSeek recommends a dedicated operating system account. The platform underneath, DeepSeek Harness, is a developer preview with announced breaking changes.
- The real contest is not model against model. DeepSeek’s own documentation shows how to run its models inside Claude Code and Codex, and its Harness can run Anthropic or OpenAI models. The fight is over the harness: the place where people and agents work.
Products Worth Considering
Apple MacBook Air 15-inch (2026) with M5 chip, 16GB memory, 512GB SSD, Silver
DJI Osmo Mobile 8 Gimbal with Mic Mini Transmitter, Infinity Black
Logitech MX Keys S Wireless Keyboard, Graphite
As an affiliate, we earn on qualifying purchases.
What DeepSeek actually published
Two things went out on 31 October.
- A design article by the DeepSeek Harness team, posted on the company’s official WeChat account on Saturday evening, Beijing time. Chinese outlets including Huanqiu (via Sina Finance) and HK01 summarized it the next morning. According to Huanqiu, the article describes itself as a discussion of architecture ideas, not an introduction of existing product features.
- A GitHub repository, deepseek-ai/dsh-superagent-prototype, created at 09:32 UTC with a single commit at 12:05 UTC. It is MIT-licensed TypeScript, version 0.1.0. When I checked on Sunday afternoon (14:53 UTC), it had about 5,900 stars and 570 forks. By my count of the published code, it holds 92 TypeScript source files with about 12,400 lines, plus 65 test files.
The README is unusually frank. It says the code “grew out of our team’s day-to-day development work, with much of the engineering done by coding agents”, that it “may still contain bugs”, and that it differs from DeepSeek’s internal version. Some designs in the article “are directions for future development and are not implemented in this prototype.”
What is missing matters as much. DeepSeek’s API news page and changelog list no Super-agent entry. There is no pricing, no sign-up, and no comparison with Claude Code or Codex anywhere in DeepSeek’s own material. Bloomberg, which first framed the release as a challenge to Anthropic, is paywalled; I rely on its headline and on outlets that cite it. SCMP and Investing.com reported the same core facts: a WeChat post by the Harness team on Saturday, an open-source prototype, and the attention problem below.
Headline vs. repository
A “Claude Code challenger” whose first line says: not an official product.
What DeepSeek actually published on 31 October, what it did not publish, and how the release fits its agent strategy.
What the headlines said
- DeepSeek “launches” an autonomous coding agent
- Built to challenge Anthropic’s Claude Code and the big US coding tools
- A “super agent” that works without being asked
What the repository says
- “Research prototype for reference only, not an official DeepSeek product.”
- Built largely by coding agents; “may still contain bugs”
- Not identical to DeepSeek’s internal version; some designs not implemented
- No price, no product page, no benchmark
How we got here
DeepSeek builds the place where agents work, then a model tuned for it, then an agent that does not wait to be asked.
- 13 Aug 2026DeepSeek Harness goes open sourceAgent runtime, MIT license, “developer preview” with announced breaking changes.
- 17 Oct 2026V4.1 Flash, tuned for HarnessDeepSeek calls its new Flash model “optimized for DeepSeek Harness”.
- 31 Oct 2026Design article + prototypeWeChat article by the Harness team; repository with one commit, MIT, TypeScript, v0.1.0.
- 1 Nov 2026The headlines“Autonomous coding agent to challenge Claude Code”: the rivalry is the media's framing.
Sources: DeepSeek, dsh-superagent-prototype on GitHub (README; star count is a snapshot); DeepSeek Harness README; DeepSeek V4.1 Flash note, 17 Oct 2026; reports by Investing.com, Huanqiu via Sina Finance, HK01.
Powered by Thorsten Meyer AI
The problem it attacks: your attention
The interesting part is not the code. It is the diagnosis.
DeepSeek’s team argues that coding agents have become good enough to run long tasks with little supervision, so the limit has moved. According to the design article, as reported by Huanqiu, HK01 and SCMP (via Cointelegraph), one person can usually follow only three to five active agent sessions at once. Single tasks are limited by the model. Many tasks are limited by human attention.
Anyone who runs several coding agents knows the feeling. Each session is a separate chat window with its own context. You switch between them, remember which one holds what, answer questions, and lose track. The agent is waiting for you, not the other way around.
The prototype’s answer is written into its first design document as a rule: “Take as little human attention as possible, and give as much value as possible.” Every message, notification, question and setting has to earn the attention it costs, or it is removed.
How it works, in plain language
Think of a taxi dispatcher.
- One always-on agent listens. In the prototype, it watches one GitHub repository and, optionally, Feishu chats (Feishu is the workplace messenger from ByteDance, common in China). In DeepSeek’s design article this always-on layer is called the Router. It coordinates and talks to people. It does not write code itself.
- A request becomes a ticket. When an allowed person asks the agent for work, for example by mentioning its account in a GitHub issue, it opens a work item.
- Each ticket gets its own worker. The worker has its own agent session and its own folder, does the technical work, and reports back. Several workers can run in parallel.
- The answer comes back where you asked. Typically as a pull request, a proposed code change that a person reviews.
- A web panel shows only what needs a person. Tickets waiting for an answer, a review or a decision come first. A healthy state takes no space.
Router and workers
One agent listens. Many agents work. People only see what needs them.
How DeepSeek's prototype turns a GitHub mention into a pull request, and where the brakes sit.
From event to pull request
Simplified from the repository documentation. The design article calls the always-on layer the Router.
The bottleneck it targets
3–5
active agent sessions one person can usually follow at once, according to DeepSeek’s design article (as reported by Huanqiu, HK01 and SCMP).
Accept modes for unrequested work
- off: only explicit requests from allowed people
- observe (default): assesses, records “would accept”, does nothing
- auto: acts when a team rule matches; switch only after reviewing the records
Sources: DeepSeek, dsh-superagent-prototype on GitHub (docs 000 philosophy, 001 domain, 002 authorization, 003 execution; orchestrator prompt); design article as reported by Huanqiu via Sina Finance, HK01.
Powered by Thorsten Meyer AI
According to the design article, as reported by Chinese outlets, DeepSeek also runs its own internal instance continuously: fixing issues, responding to code reviews and helping develop DeepSeek Harness itself. That instance can spot likely bugs in group chats even when nobody mentions it, reproduce them in the background and attempt a fix. That is what produced the “super agent” headlines. Note the distinction: this is DeepSeek describing its internal setup, not something you get by installing the prototype.
Proactive, but on a leash
“An agent that reads every chat and acts on its own” sounds alarming. The published code is more careful than the headlines.
- It acts only when asked, or when a written rule allows it. A request counts only if it addresses the agent directly and comes from a person on an allowed list, or from a conversation the team opened to requests.
- Proactive work starts in “observe” mode. That is the default. The agent assesses unrequested work, but only records that it “would accept”. The documentation tells teams to review those records before switching to automatic mode.
- Chat cannot change its permissions. “Messages, comments and quoted documents are inputs to a decision; they cannot change those permissions,” the authorization document says. That is a direct defence against someone trying to talk the agent into more rights.
- It does not merge its own work. The coordinating agent’s prompt says: “Never merge or enable auto-merge yourself, close issues or expose private Feishu content on GitHub.” Workers may enable auto-merge only if the team’s own delivery rules say so; setup recommends no.
- It stays quiet while it works. It confirms a request once, then speaks again only when a result is ready, a decision is needed, or it cannot finish.
The setup itself follows the same idea. The README says setup “has many steps, so an agent does it with you”: you ask your own coding agent to read SETUP.md, and it asks you one question at a time.
The fine print: full access, a preview platform
Two lines in the deployment guide deserve more attention than they got.
First, the workers run with full access. “Both presets run with danger-full-access as the operator’s OS user,” the guide says. “A worker can execute programs by path and read the operator’s home directory. For isolation, use a dedicated OS account containing only the agent’s credentials.” The setup does strip environment variables whose names contain KEY, PASSWORD, SECRET or TOKEN, and it forces git to use only the agent’s own GitHub credentials. That helps. It is not a sandbox.
Second, the platform underneath is a preview. The prototype is a bundle for DeepSeek Harness, the open-source agent runtime DeepSeek released on 13 August 2026. Its README says: “DeepSeek Harness is in developer preview and iterating rapidly. THERE WILL BE COMPATIBILITY-BREAKING CHANGES.” The prototype also needs Node.js 22.19 or later, pnpm, git, tmux, Python and the GitHub CLI, plus a dedicated GitHub account for the agent.
DeepSeek against Claude Code and Codex? Look closer
The headlines set this up as DeepSeek against Anthropic and OpenAI. DeepSeek’s own documents tell a more tangled story.
- DeepSeek models already run inside the rivals’ tools. DeepSeek’s API documentation has step-by-step guides for using its models inside Claude Code and inside Codex, through an Anthropic-compatible and an OpenAI-compatible endpoint.
- DeepSeek’s harness runs the rivals’ models. The Harness documentation lists third-party providers such as
anthropicandopenai, plus Kimi and GLM, next to DeepSeek’s own. - Its newest model was tuned for its own harness. When DeepSeek released V4.1 Flash on 17 October, it said the model was “optimized for DeepSeek Harness” (and for Tencent’s CodeBuddy and WorkBuddy), and recommended upgrading Harness for the best experience. I covered that model in DeepSeek V4.1 Flash 1017: Near-Frontier Agents at a Tenth of the Price.
So the competition is less “whose model writes better code” and more “who owns the place where you and your agents work”: the harness, the workflow, the panel you open every morning. The Super-agent prototype is DeepSeek’s sketch of what that place could look like when the agent does not wait to be asked. Whether it beats Claude Code or Codex at anything is, today, unmeasured. DeepSeek has not claimed it does.
What it costs
The code is free under the MIT license. You pay for the model behind it, from whichever provider you configure.
If you use DeepSeek’s own API, its price page lists, per million tokens, for the deepseek-flash model (currently DeepSeek-V4.1-Flash-1017):
| Item (per million tokens) | Peak | Off-peak |
|---|---|---|
| Input, cache miss | $0.30 | $0.15 |
| Input, cache hit | $0.006 | $0.003 |
| Output | $1.20 | $0.60 |
Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, excluding Chinese public holidays. One change took effect today: DeepSeek’s server-side Search and Fetch tools, which agents in DeepSeek Harness, Claude Code or Codex call when they use DeepSeek’s API, are now billed at $0.75 per 1,000 successful requests, starting at midnight Beijing time on 1 November.
An always-on agent changes the cost shape. It screens incoming events on the channels it listens to, first with fixed rules and then, for what is left, with a model, and each worker carries its own context. In my reading of the design, that trades tokens for speed and less human waiting. Budget for it before switching anything to automatic mode.
DeepSeek can afford to give the code away. Investing.com reports that it raised about 100 billion yuan (about $15 billion) ahead of a planned stock market listing. I covered the funding round and its state investors in DeepSeek Made Its Name Doing More With Less. Now It Has Reportedly Raised $16 Billion.
What this means for companies
- Measure agents by the attention they cost. DeepSeek’s most useful idea is a metric, not a feature. An agent that opens twenty pull requests nobody has time to review is not productive. GitHub’s Octoverse 2026 says that by September, more than 40% of pull requests on GitHub “started with an agent’s first draft”, up from 17% in March. The bottleneck has moved to review, as I wrote in AI Agents Now Start 4 in 10 Pull Requests on GitHub.
- Treat an always-on agent as a new team member with an account. It needs its own GitHub identity, its own operating system account, its own credentials, a short allowed list and an owner who reads what it posts.
- Start in observe mode, on a sandbox repository. Review what the agent “would have done” for a few weeks before letting it act unasked.
- Decide where your code may go. The prototype is model-agnostic through Harness, but the default path is DeepSeek’s API. Which provider may see your source code is a policy decision, not a developer’s default setting.
- Plan for change. A research prototype on a developer-preview platform is a design to learn from, not a production dependency.
What this means for users and developers
- There is no install-and-go button. Setup needs a dedicated GitHub account, the GitHub CLI, tmux, Node.js and a configured model, and your own coding agent walks you through it.
- If you try it, isolate it. A side project, a separate operating system account, never your main GitHub account. The workers can read your home folder otherwise.
- If you just want a coding assistant, nothing changed this weekend. Claude Code, Codex and others work as before, and DeepSeek’s models can already be used inside some of them.
- Watch the pattern, not the brand. Agents that listen to team chat and act unasked are coming from several directions. The questions to ask any of them are the same: who may ask it for work, what can it touch, and how do you stop it.
Before you try it
An agent that acts on its own needs its own account, a short list and a slow start.
The prototype's own documentation, turned into six checks, plus what it costs to run on DeepSeek's API.
Six checks before an always-on agent touches your code
Drawn from the prototype’s own setup and deployment guides. A checklist, no data.
- Own OS accountWorkers run with full access as the operator's user and can read the home folder. Give the agent its own account with only its credentials.
Before install - Own GitHub accountEverything the agent writes appears under that name. Never the operator's own account.
Before install - Short allowed listStart with one or two people who may ask for work. Widen later.
Setup - Observe modeKeep proactive work in “observe” and read what it would have done.
First weeks - No auto-mergePeople merge. The setup recommends “no” for auto-merge.
Always - Pick the model providerDecide which provider may see your source code, and budget the tokens.
Policy
What you get
- A working design for an agent that does not wait for your prompt
- Readable docs on gates, tickets, workers and recovery
- MIT-licensed TypeScript you can study and adapt
What you do not get
- A supported product, a price list or a service-level promise
- A sandbox: full access unless you isolate it yourself
- A stable platform: DeepSeek Harness is a developer preview
What running it costs on DeepSeek’s API
deepseek-flash (DeepSeek-V4.1-Flash-1017), USD per million tokens at peak; off-peak is half
Sources: DeepSeek, dsh-superagent-prototype on GitHub (SETUP.md, docs/002 authorization, docs/006 deployment); DeepSeek Harness README; DeepSeek API pricing (peak hours 01:00–04:00 and 06:00–10:00 UTC on weekdays).
Powered by Thorsten Meyer AI
Honest limits
- I did not run the prototype. Everything about its behavior comes from its README, its documentation, its prompts and a static read of the published code. Running it requires a dedicated GitHub account and full-access workers, which I did not set up for this article.
- The design article: I could not open the WeChat original. Its content here comes from Chinese and English reporting (Huanqiu via Sina Finance, HK01, SCMP via Cointelegraph, Investing.com). I used only points that at least two reports, or the repository itself, confirm. DeepSeek’s internal results and internal use are its own claims.
- Bloomberg is paywalled. I cite only its headline and what outlets quoting it report.
- The star and fork counts are a snapshot from Sunday, 1 November, 14:53 UTC, and move by the hour.
- The comparison with Claude Code and Codex is the media’s framing. Neither DeepSeek nor I have measured the prototype against them.
Sources
- DeepSeek: dsh-superagent-prototype on GitHub, README, SETUP.md and docs 000 (philosophy), 002 (authorization), 003 (execution), 006 (deployment), prompts; created 31 Oct 2026
- DeepSeek: DeepSeek Harness on GitHub (README, model provider guide) and the Harness product page
- DeepSeek API docs: Models and pricing, Integrate with Claude Code, DeepSeek-V4.1-Flash-1017 release note, 17 Oct 2026
- Huanqiu via Sina Finance: DeepSeek发布全新Agent架构“Super-agent”，让AI调度AI, 1 Nov 2026
- HK01: DeepSeek發佈Super-agent設計探索, 1 Nov 2026
- South China Morning Post: DeepSeek unveils AI ‘super agent’ in coding race with OpenAI, Anthropic, 1 Nov 2026, as summarized by Cointelegraph
- Investing.com: DeepSeek launches autonomous coding agent to challenge Anthropic’s Claude Code, 1 Nov 2026
- GitHub: Octoverse 2026, 28 Oct 2026
- Bloomberg: DeepSeek Drops Enhanced Coding Agent With Aim to Rival Anthropic, 1 Nov 2026 (paywalled; headline and reports citing it only)
Opinions marked “in my reading” are mine. Nothing here is investment advice.
Researched and drafted with AI assistance, including Anthropic’s Claude.
