What DeepSeek published on October 31
The DeepSeek Harness team put dsh-superagent on GitHub. The README calls it a research prototype for reference only, and not an official DeepSeek product. package.json is version 0.1.0 under the MIT licence, and the package is marked private. The README says the published code is not the internal version, and not the implementation in the design article Exploring Super-agent Design: Towards a More Autonomous Agent System. Some ideas in that article are not built here.
It is a bundle for DeepSeek Harness. One deployment serves one GitHub repository, and it can also listen to Feishu. When someone the team has allowed asks for work, it opens a work item, does that work in its own agent session and its own directory, and answers in the original conversation, for example with a pull request. A web panel lists what still needs a person.
The README states the aim as taking as little human attention as possible and giving as much value as possible. These are DeepSeek's descriptions of its own prototype. This page is not a CuratedAI run of it, and the prototype is not a directory listing.
It stays quiet unless a person or a written rule starts it
The authorization doc, 002 Authorization, is the source for when work starts. An explicit request from a person on the allowed list works in every mode. Work the agent only infers, such as a message that does not address it, depends on the accept mode. The default is observe. In that mode a positive assessment is recorded as "would accept" and nothing is accepted. auto is the switch that lets a matching team rule start that work. off admits no inferred work.
The README says the agent acts only when asked, or when a rule the team wrote allows it, and that it speaks only when someone must act or can use a result. A bot on the source's bot list can only produce an inferred candidate. It cannot make an explicit request or stop work. DeepSeek says to start in observe and move to auto only after reading the recorded decisions.
The package requires Node.js 22.19, or Node.js 24 or newer, and DeepSeek Harness 0.2.1-alpha.2 or newer. Setup is a long checklist. The README says to have a coding agent read SETUP.md with you. pnpm check runs the typecheck, the tests, and the build.
Match the job to a read, a trial, or a tool you already open
The prototype is a way to study an agent that sits on a repository. It is a poor plan for any job you need finished today.
observe. Read the "would accept" decisions before anyone sets it to auto. One deployment is one repository. Feishu is optional, and an unlisted person who addresses the agent is ignored unless the team changes that cell.Where the coverage is easy to overread
Bloomberg reported that the agent can watch developer chats and suggest fixes without being summoned, and that DeepSeek aims to rival Anthropic. The README is narrower. Work that nobody asked for depends on a rule the team enabled, and on leaving observe mode. DeepSeek also says the public code can still contain bugs, and that its style is uneven because much of the engineering was done by coding agents.
Star counts and fundraising figures in the same news cycle are not in the repository. A news story about the prototype is not a test of it. The same day's briefing covers a reported cut to OpenAI's funding target, Aramco's stake in HUMAIN, a gene-therapy experiment on a live mouse, and a tampered Claude Code extension on Cursor. Those are separate stories. Read them in the November 1 briefing.
What to do this week
If you run agents against a repository you control, read the philosophy doc and the authorization doc before you install anything. If you need a coding agent today, keep the one a person starts. For the rest of the day, read the November 1 briefing. Reviewed November 2, 2026, from the dsh-superagent README, package.json, and the authorization doc.
Questions
- When should a team try DeepSeek's Super-Agent, and when should it stay on the coding agent it already runs?
- The DeepSeek Harness team put dsh-superagent on GitHub.