DeepSeek has published the code of Super-agent, a prototype autonomous coding agent designed to compete with Anthropic's Claude Code, Bloomberg reported on Sunday. The code was published on GitHub after the company's Harness team described the agent in a WeChat post on October 31. The release is distributed under the MIT licence, which allows anyone to use and modify the code, and this matters because companies can test background bug-fixing without buying a commercial coding tool.
How Super-agent works in background mode
The agent can monitor developers' group chats and identify possible bugs even when nobody mentions it directly, according to the post cited by Chinese outlet Yicai. Once it spots a potential problem, it attempts to reproduce the issue and prepare a fix in the background. DeepSeek already uses Super-agent internally to fix bugs, respond to code reviews, and improve its own tools. The stated design goal is to take up as little human attention as possible during that work.
The prototype runs on DeepSeek Harness, the company's open-source software that wraps its AI models so agents can handle longer and more complex tasks. It works with GitHub, with optional support for Feishu, the workplace messaging app owned by ByteDance. According to project documentation, the agent starts work only when an approved person requests it or when a team-defined rule permits it. That permission model separates passive chat observation from the authority to act on code.
The release fits a contest with coding tools from US labs. DeepSeek formed the Harness team earlier this year specifically to build agents able to compete in that segment, Bloomberg reported. The Super-agent announcement followed a reported $15bn funding round in October with Tencent and battery maker CATL among investors, ahead of a possible stock market listing. DeepSeek has not confirmed that round, and its GitHub page calls Super-agent a research prototype for reference only, not an official product.
What autonomous bug-fixing means for business
For companies that maintain software, the practical effect is a lower entry barrier to testing an agent that triages chats, reproduces defects, and drafts fixes. Small teams can run the MIT-licensed code against GitHub repositories and add Feishu if they use that messenger, without changing vendors. Larger organisations can assign it to code-review queues and internal tooling maintenance, where background work reduces waiting time between a reported symptom and a reproducible case.
The same design creates verification costs that buyers should price in advance. The GitHub page advises developers to review and adapt the code before use, so deployment requires code audit, access controls, and a clear list of approved requesters and activation rules. The news by itself does not prove accuracy of bug detection, quality of generated fixes, or safe behaviour in production repositories. Questions to the team should cover false positives in chat, reproduction success rate, write permissions, and logging of agent actions.
The marker to watch is whether DeepSeek moves Super-agent from reference prototype to supported product with documented deployments outside its own engineering. A release schedule, enterprise support terms, or confirmed use by an external company would show the approach survives real repositories. Until then, the business value stays limited to controlled pilots with human review of every proposed change.
