DeepSeek open-sources Super-Agent prototype for autonomous coding
DeepSeek has released an open-source research prototype designed to monitor developer activity around the clock and draft code fixes, positioning the project as a move toward autonomous software development.
DeepSeek has open-sourced an autonomous coding agent prototype under the MIT license in a repository named deepseek-ai/dsh-superagent-prototype, subtitled 'Towards a More Autonomous Agent System'. According to reporting by Bloomberg, the release marks an effort by the Chinese artificial intelligence company to challenge Anthropic's Claude Code tool.[1][2][3]
The prototype was created by the team behind DeepSeek Harness, the company's internal agent framework. The system stays online 24/7 to monitor a designated GitHub repository as well as developer group chats, such as Feishu. When an approved user requests assistance in an issue or pull request, the agent initiates a work item, prepares an initial pass, drafts code fixes, and submits the changes for human review.[1][2][4]
DeepSeek emphasizes that the codebase represents experimental research rather than a commercial product. In an accompanying design write-up titled 'Super Agent: Proactive Is All You Need?', harness engineer Yujia Qiao detailed an architecture consisting of an always-on Router paired with multiple parallel Workers. The system can be triggered by group-chat messages, emails, calendar reminders, GitHub comments, and continuous integration test results. The design seeks to reduce the strain on human oversight, addressing a practical ceiling where individual developers can generally manage only three to five agent sessions simultaneously.[1]
Key facts
- DeepSeek released the deepseek-ai/dsh-superagent-prototype repository on GitHub under the MIT license.
- The prototype was developed by the team behind DeepSeek Harness and grew out of their internal development work.
- The agent monitors a GitHub repository and optionally Feishu chats, drafting fixes for issues and pull requests before submitting them for human review.
- The system operates 24/7 using an always-on Router and parallel Workers triggered by chats, emails, calendar reminders, GitHub comments, and CI test results.
- DeepSeek engineer Yujia Qiao authored the project design write-up, 'Super Agent: Proactive Is All You Need?'.
- The project aims to minimize human intervention, citing an estimated limit of three to five agent sessions a person can effectively manage at once.
- Bloomberg reported the release as a move by DeepSeek to compete with Anthropic's Claude Code.
Sources · 4 sources
- AG
AGTP@AGTPinsightsPost on X ·
DeepSeek just open-sourced dsh-superagent-prototype, an agent that works your GitHub repos. Here's what you need to know. It comes from the team behind DeepSeek Harness, the company's agent tool. It watches one GitHub repo, and optionally Feishu chats. When someone on an allowed list asks for something in an issue or PR, it opens a work item, takes a first pass, drafts a fix, and hands the result back for human review. The code is MIT licensed. DeepSeek also published a design write-up, "Super Agent: Proactive Is All You Need?", by Yujia Qiao, an engineer on the harness. It first spread in China, and community translations into English followed within hours. The team says an official English version is coming soon. The problem they describe is attention. People can only juggle about 3 to 5 agent sessions at once, so human attention becomes the ceiling. Their stated goal is to take as little human attention as possible and give as much value as possible. The design has an always-on Router and many parallel Workers. Triggers include group-chat messages, emails, calendar reminders, GitHub comments and CI test results. The agent stays online 24/7 and picks up work on its own. DeepSeek calls it research code, not a product. It grew out of the team's own day-to-day work on the harness. Community reaction so far: - Supporters see a possible "software factory" if it is set up well. - Other builders say their own always-on agent systems landed on similar designs. - Critics say the framing is about replacing slow humans with agents running many sessions at once. - One prominent DeepSeek watcher claims that by late V4, DeepSeek's internal agent swarms reached the utility of Anthropic's Mythos Preview. They say it took swarms of 200 to 1000 agents, with a coordinator launching each next swarm from the last one's findings. They also claim the baseline is five swarms of 1000 agents per night. These are unconfirmed claims, not DeepSeek statements. Key numbers: - 3 to 5: agent sessions a person can manage at once - 24/7: how long the agent stays online - MIT: license of the prototype - 200 to 1000: agents per swarm in the unconfirmed claims The repo is deepseek-ai/dsh-superagent-prototype on GitHub, subtitled "Towards a More Autonomous Agent System".
Open source - TE
TechmemeArticle ·
DeepSeek open sources Super-Agent, a coding agent research prototype it says proactively monitors developer group chats for coding issues and suggests solutions (Debby Wu/Bloomberg) Debby Wu / Bloomberg : DeepSeek open sources Super-Agent, a coding agent research prototype it says proactively monitors developer group chats for coding issues and suggests solutions — DeepSeek has released a prototype of a more autonomous coding agent as it seeks to challenge Anthropic PBC's popular Claude Code.
Open source - BL
Bloomberg@businessPost on X ·
DeepSeek has released a prototype of a more autonomous coding agent as it seeks to challenge Anthropic’s popular Claude Code https://t.co/7VFHRdFuHF
Open source - CO
Cointelegraph@CointelegraphPost on X ·
🤖 NEW: DeepSeek has released “Super-Agent,” an open-source coding agent prototype that monitors developer group chats and suggests solutions without prompting, aiming to rival Anthropic’s Claude Code. https://t.co/YHjm32zkSK
Open source