DeepSeek releases open-source 'Super-agent' in direct challenge to Anthropic's Claude Code

DeepSeek, the Hangzhou-based artificial intelligence company, released an open-source prototype of an autonomous coding agent on Oct. 31, posting it to GitHub under the name Super-agent as part of a broader push to compete with Anthropic's Claude Code. The release was announced through a post from DeepSeek's internal "Harness" team on WeChat, with the underlying code and documentation published directly by DeepSeek.

Unlike a conventional assistant that waits for a prompt, Super-agent is designed to act on its own. According to the project's GitHub repository, the tool monitors a single GitHub project and, optionally, the Chinese workplace messaging app Feishu, watching group chats among developers for mentions of coding problems. When it identifies an authorized request, it opens a work item, carries out the task in its own isolated agent session, and reports back in the original conversation, typically with a pull request. A web dashboard flags items that still need human attention. DeepSeek states the project's goal directly in its documentation.

"Take as little human attention as possible, and give as much value as possible."

The repository had drawn roughly 6,900 stars and nearly 700 forks within a day of wider coverage, by GitHub's count, a sign of fast developer interest even though DeepSeek is explicit that the release is not a finished product. The README describes it as "a research prototype for reference only," built largely with the help of coding agents themselves, and warns that it "is not production-ready," may contain bugs and differs from both DeepSeek's internal version and from an accompanying design essay, titled "Exploring Super-agent Design: Towards a More Autonomous Agent System," that lays out features not yet implemented in the public code.

From models to "harnesses"

The release is the latest step in a strategy DeepSeek began building earlier this year, when it formed a dedicated team to compete in what the industry calls the "harness" layer: the surrounding software that gives a language model tools, memory and the ability to plan multi-step tasks, rather than the model itself. DeepSeek's own framing, used in its internal materials, is "Model + Harness = Agent." Super-agent is packaged as an add-on to DeepSeek Harness, a broader open-source agent runtime that DeepSeek's own site describes as being "in public preview worldwide" and licensed, like Super-agent, under the permissive MIT license.

The target is explicit. DeepSeek's effort is aimed squarely at Claude Code, Anthropic's flagship coding-agent product, which has become one of the fastest-growing parts of Anthropic's business. Anthropic disclosed in its own funding announcement that Claude Code's run-rate revenue has grown to over $2.5 billion, having more than doubled since the start of 2026, with weekly active users also doubling and enterprise accounts now making up more than half of that revenue. That scale is the prize DeepSeek, OpenAI's Codex and other rivals are chasing as model performance converges and companies compete instead on how well an agent integrates into a developer's actual workflow.

A rivalry with a security backdrop

The launch lands amid open friction between DeepSeek and Anthropic. In February, Anthropic published a post accusing DeepSeek, along with two other Chinese labs, Moonshot AI and MiniMax, of running what it called an industrial-scale campaign to extract Claude's capabilities through roughly 24,000 fraudulent accounts and more than 16 million exchanges, a technique known as distillation. In that same post, "Detecting and preventing distillation attacks," Anthropic said DeepSeek alone was responsible for more than 150,000 of those exchanges and used the episode to argue for tighter export controls on the advanced chips needed to run such attacks at scale.

DeepSeek's new agent also arrives just over a week after the U.S. government moved against the company on security grounds. On Oct. 20, a federal directive ordered civilian agencies to remove DeepSeek's AI technology from government systems, citing security concerns, a move TechCrunch reported days before DeepSeek was said to have closed a roughly $15 billion funding round, backed by investors including battery maker CATL and Tencent, ahead of a planned initial public offering in China as early as next year. The combination means a tool designed to act with minimal supervision inside developers' code repositories and chat channels is coming from a company currently barred from U.S. government networks, a juxtaposition likely to draw scrutiny from security teams even among developers simply curious about the open-source code.

Who is affected, and what comes next

For now, the practical audience is narrow: developers willing to install research-grade, MIT-licensed code into their own GitHub projects, with DeepSeek urging anyone who does so to "review and adapt it" first rather than deploy it as-is. The company has not published benchmark comparisons against Claude Code, OpenAI's Codex or other agent harnesses, and neither Anthropic nor OpenAI had publicly responded to the release as of Sunday afternoon. Multiple outlets picked up the story Sunday — South Korea's Digital Today credited Bloomberg's original report directly, while Investing.com ran its own version of the story — but none added independent reaction from developers or analysts beyond DeepSeek's own materials.

What happens next will depend on whether DeepSeek narrows the gap between the published prototype and the fuller "Super-agent" vision laid out in its design essay, and on how quickly rivals respond. Anthropic, OpenAI and startups building their own agent harnesses are all racing to lock developers into their own tools before habits set, the same dynamic that is pushing DeepSeek to give its version away for free under an open license. DeepSeek has not said when, or whether, Super-agent will graduate from research prototype to a supported product.