DeepSeek publishes a Super-agent prototype and says the code is not the system it runs
Bloomberg’s account, in The Straits Times on Sunday morning, says a Saturday WeChat post described an agent that can watch developer chats and suggest fixes without being asked. The GitHub repository, dated 31 October, calls the code a research prototype, not an official product, and not the same as the internal system or the essay.
Feature3 minIn-depth
James Okonkwo3 min read
- DeepSeek
- Agents
- Claude Code
- Harness

Photo: Reuters, via The Straits TimesSource page
DeepSeek has published a research prototype of a coding agent it calls a Super-agent, and the repository says the code is not the system the company runs. Bloomberg’s account, carried by The Straits Times in a story timestamped 12:55 p.m. in Singapore on Sunday, 5:55 a.m. in Amsterdam, says the Harness team described the tool in a WeChat post on Saturday, and that the agent can watch developer group chats and suggest fixes without being asked. The same story takes a line from a GitHub manual: take as little human attention as possible, and give as much value as possible. The Straits Times attributes the piece to Bloomberg. The page does not print a Bloomberg byline.
The repository, deepseek-ai/dsh-superagent-prototype, is dated 31 October on GitHub. The readme’s first paragraph says it is a research prototype for reference only, not an official DeepSeek product, that much of the engineering was done by coding agents, and that the code may still contain bugs. It also says the published version was adapted for release, so it is not identical to the internal version or to the implementation described in an essay, Exploring Super-agent Design: Towards a More Autonomous Agent System. Some designs in that essay, the readme says, are directions for future development and are not in this prototype. The license file is the MIT license, copyright 2026, DeepSeek Harness Team.
What the published code says it will do
The readme describes dsh-superagent as a proactive coding agent for one GitHub repository, packaged as a bundle for DeepSeek Harness. It listens to GitHub and, optionally, Feishu. When someone the team has allowed asks for work, it opens a work item, works in its own session and directory, and answers in the original conversation, for example with a pull request. A web panel shows what still needs a person. Setup, the readme says, is long enough that the intended path is to hand SETUP.md to a coding agent and answer one short question at a time.
A philosophy note in the same repository is the rule under the “without being asked” line in the wire story. The agent acts only when asked, or when a written rule allows it. A request counts only when it addresses the agent and comes from a person on an allowed list, or from anyone in a conversation the team opened to requests. Work nobody asked for needs a rule the team wrote, and it starts in a mode that only records what the agent would have done. The note also says the agent speaks only when a person must act or can use a result: it confirms a request once, stays quiet while it works, and writes again when a result is ready, when it needs a decision, or when it cannot finish.
Take as little human attention as possible, and give as much value as possible.
— DeepSeek Harness team, dsh-superagent readme and philosophy note
What a Saturday brief says the essay claimed
Shanghai Metals Market published an English brief at 23:01 GMT+8 on Saturday, 4:01 p.m. in Amsterdam. It says the Harness team that day posted “Super-agent Design Exploration: Toward More Autonomous Agent Systems,” and that the essay argues mainstream agents work in isolated sessions, so one task is capped by the model and several tasks are capped by human attention. The brief says the team’s answer is a dual-layer architecture, event-driven asynchronous interaction across channels, a split between storage and compute, and an external watchdog, with session management moved onto the system. It says an internal instance is already running on bug reproduction, code review, and work on Harness itself. The brief does not quote the WeChat post. The readme’s warning covers that gap: the code on GitHub is not the essay, and it is not the internal instance.
Bloomberg’s account, in the Straits Times version, also says DeepSeek created the Harness team earlier in 2026 to build agents that can take on coding work offered by labs including Anthropic, and it defines a harness as the scaffolding around a model. The sentence a developer can check against the repository is narrower than the wire story’s lead. The prototype acts when asked, or when a rule the team wrote allows it, and unsolicited work starts as a record of what the agent would have done.


