Key takeaways

  • DeepSeek released an MIT-licensed research prototype of a proactive coding agent for a single GitHub repository.
  • It acts on explicit requests from allowed people, and on inferred ones only when a team-written rule matches.
  • DeepSeek calls it a prototype for reference only and has published no benchmarks.

DeepSeek has open sourced a research prototype of a "super-agent", a coding agent that watches a GitHub repository and picks up work when a person asks or a team-written rule allows it. The code for the DeepSeek super-agent prototype is under the MIT license, and Bloomberg reported the release on November 1.

It matters because most coding agents wait for a prompt. This one is built to sit in a team's chat and issue tracker and decide for itself which requests are worth acting on. It also lands as a direct challenge to Anthropic's Claude Code, which is how Bloomberg and the South China Morning Post framed it.

What the super-agent does

The README describes a proactive coding agent for a single GitHub repository. It takes work requests from allowed users, opens a work item, does the work in its own agent session and directory, and replies in the original conversation, for example with a pull request. A web panel lists the items that need a person.

DeepSeek states the goal as using as little human attention as possible while producing as much value as possible. In practice the agent speaks only when someone has to act or can use a result. It listens to GitHub events and, optionally, to Feishu, the chat app from ByteDance.

How it decides what to do

Requests from a listed person are explicit, and the agent acts on them. Anything else is "inferred", for example a bot posting in a watched channel or an unlisted person following up on a ticket. An inferred request is judged against rules the team writes, and the agent accepts it only if at least one rule matches.

There are three accept modes:

  • off: nothing inferred is admitted.
  • observe: the default. The agent records what it would have accepted and accepts nothing.
  • auto: a positive assessment of an inferred request is accepted.

DeepSeek's docs suggest starting in observe, reviewing the recorded decisions, then moving to auto. Bots can never make explicit requests. Messages and quoted documents are treated as input to a decision and cannot change permissions.

Setup and warnings

The setup guide lists Node.js 22.19 or newer, pnpm, git, tmux, Python and the GitHub CLI. It also needs a model provider configured in DeepSeek Harness, the tool the prototype is packaged for, and a separate GitHub account for the agent. Auto-merge is off by default. The page does not name which model the agent runs on.

The README calls the project a "research prototype for reference only, not an official DeepSeek product". It says the code may contain bugs, differs from DeepSeek's internal version and from the design article it accompanies, and implements only some of that article's ideas. At the time of reading, the repository showed one commit and about 4,900 stars.

What we don't know yet

DeepSeek has not published benchmarks for the agent, so there is no way to say how well it works against Claude Code or Codex. The design is the news, not a score. Anyone who tries it should start in observe mode on a throwaway repository, since an agent that acts on events from the outside world is exposed to prompt injection, which means text in an issue or message that tries to steer the agent.

For the model side of DeepSeek's lineup, see our explainer on DeepSeek V4.1 Flash.

Sources

  1. dsh-superagent-prototypeDeepSeek (GitHub)
  2. Authorization modelDeepSeek (GitHub)
  3. Setup guideDeepSeek (GitHub)
  4. DeepSeek unveils AI super agent in coding race with OpenAI, AnthropicSouth China Morning Post
  5. DeepSeek open sources Super-Agent as a research prototypeTechmeme (Bloomberg)