DeepSeek open-sources a coding agent that acts without being asked

DeepSeek’s Harness team has open-sourced a prototype of a proactive coding agent it calls a “super agent”, The Information reported. The code is on GitHub under an MIT license, and DeepSeek labels it a research prototype, not a product.

What the prototype does and how to run it

The repository, deepseek-ai/dsh-superagent-prototype, was published on October 31. According to its README, the agent is built for a single GitHub repository and is packaged as a bundle for DeepSeek Harness, the company’s agent runtime. It listens to GitHub activity and, optionally, to Feishu messages.

When a person on an allow list asks for something, the agent opens a work item and handles it in its own agent session and working directory. It then answers in the conversation where the request came from, for example with a pull request. A web panel lists the items that need a human decision.

The README describes the design goal as using as little human attention as possible while still delivering value. The agent acts when asked, or when a rule written by the team permits it, and it speaks up only when someone has to act or can use a result. Bloomberg, as summarized by Techmeme, described it as able to monitor group chats for coding problems and suggest solutions.

Setup has two routes. The README suggests pointing a coding agent at SETUP.md, which asks one question at a time, shows the full plan once, then installs and configures everything. A manual route is documented in docs/006-deployment.md. Developers can run pnpm check to execute the typecheck, tests and build. New input channels can be added as “Sources”, with a cookbook in docs/008-cookbook.md.

A prototype, with caveats DeepSeek states itself

The README is blunt about the limits. It warns of possible bugs and uneven engineering, and tells users to review and adapt the code before relying on it. It also says the release differs from DeepSeek’s internal version and from the system described in a companion article, “Exploring Super-agent Design: Towards a More Autonomous Agent System”. Some designs in that article are future directions and are not implemented here.

The README names no supported models and reports no benchmarks, so there is no way yet to compare it with Claude Code or Codex on output quality. The claim that it can compete with those tools comes from press framing, not from DeepSeek’s documentation.

From a tool you prompt to one that waits for work

The technical shift is in how work arrives. Tools such as Claude Code and Codex are mostly started by a developer in a terminal or editor. This prototype sits inside the channels where a team already talks, treats a message or issue as a task, and reports back without a developer supervising each step. The same README puts the human in the loop only for authorization and for items that land in the panel.

That design adds a security question that a prompt-driven tool avoids: who is allowed to trigger the agent, and what it may do on its own. The README lists authorization as its own numbered document, which suggests DeepSeek treats it as a core concern, though the text of that document was not part of what was reviewed here.

A fast-moving open-source track record

The release builds on DeepSeek Harness, which coverage from secondary outlets describes as an open-source agent runtime released as a developer preview in mid-August under the MIT license. Those reports described it as raw and closer to a scaffold than a finished product, and the official repository was not checked for those details.

On October 20, DeepSeek published DeepInfer, the inference stack behind its own services, under Apache 2.0. That release left out parts such as a KV cache system and load balancing across replicas. The super agent follows the same pattern of publishing real internal tooling in a deliberately incomplete form.

Developer interest is visible. At the time of writing, the repository showed about 9.4k stars and 956 forks. Stars measure attention, not adoption, and say nothing about how well the agent performs.

For now, anyone can read the code, run it against one repository and judge it. What is missing is evidence: no benchmarks, no model list, and no statement on when, or whether, a supported product will follow. DeepSeek has not publicly said.