DeepSeek open-sources Super-agent, a coding agent aimed at Claude Code

What it's about

DeepSeek's Harness team open-sourced a coding agent called Super-agent on October 31, under the MIT license. It connects to a GitHub repository and can also monitor Feishu group chats (Feishu is a team messenger used mainly in China). Authorized team members send requests through either channel. The agent does the work and reports back where the request was made, if needed as a pull request (a proposed change to code that others review and merge).

Outlets such as Bloomberg reportedly framed the release as an attack on Anthropic's Claude Code, shortly before Anthropic's planned IPO. DeepSeek itself, by contrast, describes the project as research software, not an official product and not production-ready. The public version also differs from the system the team uses internally.

How Super-agent works

The agent is built for exactly one repository. An authorized requester posts a task via GitHub or Feishu. The agent turns it into a work item (a "ticket") with its own session and working directory. Progress or results show up in the original conversation. A local web panel shows cases where a human needs to make a decision.

There are two internal roles:

• Router: a long-running session that takes in requests and holds shared context. It can accept new requests while others are still in progress, so tasks overlap.
• Worker: separate sessions that do the actual work. They also track comments on their pull requests.

Teams can allow work that follows written rules. The default, however, is observation mode: the system records what it would do on its own but doesn't execute it. Operators can review these proposals and only then enable automatic acceptance. This lets you raise permissions step by step and limit unwanted activity.

Setup and requirements

Setup is not trivial. According to Blockonomi, you need:

• a dedicated GitHub user with write access to the repository (github.com is supported, GitHub Enterprise Server is not),
• Node.js 22.19 or newer (or 24+), pnpm, Git 2.32+, Python 3.8+ and tmux,
• GitHub CLI 2.48+ and a supported version of DeepSeek Harness,
• a configured model provider,
• for Feishu, optionally a bot and lark-cli.

A separate account for the agent handles its GitHub identity and signing.

The security risk

The most sensitive point is process access. The router and workers run with full access as the operator's OS user. A worker can run programs by path and read that user's home directory. The GitHub credential wrappers select the agent's account, but they don't isolate the process. So the agent's GitHub identity does not limit its local permissions.

DeepSeek therefore recommends a dedicated OS account that holds only the agent's credentials. Because the agent can access team chats and repository activity, teams should also carefully review who is allowed to make requests and which repository permissions apply, and start with narrowly scoped tasks. The README also warns of possible bugs and inconsistent engineering practices; you should review and adapt the code before using it in a different environment.

Numbers and reaction

On November 1, GitHub showed 6.6k stars and 652 forks. According to Blockonomi, that shows early interest, but not performance. The repository contains no comparative benchmark for accuracy, speed or reliability, and no evidence that Super-agent can keep up with Claude Code.

Takeaway

Super-agent is less a new model than a working framework. It shows how an agent can be embedded in the daily routine of a team built around issues, chats and code review, with a sensible, cautious default (observation mode). For developer teams interested in agent orchestration, it's an intriguing toolkit, but it is deliberately not meant for production use yet. Based on the current facts, it doesn't qualify as a "Claude Code killer" as long as performance comparisons are missing. The broad system permissions are the critical issue: without isolation, the risk to your credentials is real. It remains to be seen how the public version evolves relative to the internal one. Note: the other sources provided on DeepSeek V3.1 and the DeepInfer inference framework cover different stories and were not used here.