DeepSeek Harness Open Sources a Super Agent Prototype That Picks Up Coding Work From GitHub on Its Own
Context
It shows a major Chinese lab testing coding agents that decide for themselves which work to pick up, rather than waiting for a person to prompt them.
DeepSeek Harness is an X account tied to DeepSeek's agent tooling, and its post was reposted by the main DeepSeek account. The post is an X Article titled "Exploring Super-agent Design: Towards a More Autonomous Agent System", published alongside a code repository under the deepseek-ai organization on GitHub. Polymarket framed the release as a rival to Anthropic's Claude Code.
Deep dive
DeepSeek Harness published an X Article on Nov 1 describing what it calls a super agent: a coding agent meant to work more autonomously than today's chat based assistants. Code for a prototype sits in the deepseek-ai organization on GitHub.
According to the repository's documentation, the prototype is a research project and not an official DeepSeek product. It is packaged as a bundle for DeepSeek Harness and works on a single GitHub repository. Instead of waiting for a prompt, it listens for activity on GitHub, and optionally on the Feishu chat app, and decides which events need work.
The documentation describes a long running component called the Router that evaluates each trigger. Work it accepts becomes a Ticket with its own worker session and working directory, and the agent answers in the conversation where the request came from, for example by opening a pull request. A web panel shows the items that need a person to look at them.
The repository warns that workers run commands without asking for approval, as the operating system user that runs the prototype, so it should be run with care. It also says some designs described in the article are not implemented in the prototype yet.
Polymarket reported the release to its followers as DeepSeek launching an open source coding agent designed to rival Claude Code. Performance figures circulating on X about the agent have not been confirmed by DeepSeek.
What to expect next
Watch whether DeepSeek turns the prototype into a supported product or folds the ideas into its own models and apps, and whether it publishes measured results for the agent. Developers trying it will report how well the Router chooses work and how safe it is to let workers run commands without approval. Expect comparisons with Claude Code, Codex and other coding agents that are also moving toward working on their own in the background.