DeepSeek Super Agent: can you use it, and should you trust it?

The DeepSeek Super Agent is a proactive coding agent: software wrapped around an AI model that lets it take on whole jobs, such as reproducing a bug, writing the fix and replying to code review, instead of answering one question at a time. The code sits on GitHub under the name dsh-superagent with an MIT license, which means anyone can use, change and ship it, including in commercial work.

It comes from a DeepSeek team that builds agent tools, set up earlier this year to compete with coding products from US labs. The team explained the design in a long WeChat post on Saturday. A Bloomberg report framed the release as a direct shot at Anthropic’s Claude Code.

It’s an add-on for DeepSeek’s open agent framework, first previewed in August. Each installation covers one GitHub repository, with Feishu, the workplace chat app made by ByteDance, as an optional second place to listen.

The big idea is that human attention, not AI smarts, is now the bottleneck. DeepSeek’s team says one person can follow only three to five active AI sessions at a time. So instead of making you juggle sessions, the Super Agent splits the job in two.

• The router is the part that talks to people. It keeps context across tasks and hands out work.
• Workers each own one task in their own folder, usually a separate git worktree. They report to the router rather than chatting with people, though on GitHub they open the pull request and handle review comments themselves.
• A web panel lists anything that needs a human decision.

By default there are three worker slots, plus a fourth held back so a person’s follow-up never gets stuck in line behind new work. Every accepted job becomes a ticket with a locked record of who asked, which rules matched and how far the work may reach. Anything bigger needs a fresh ticket.

The rules of conduct are strict on purpose. The agent confirms a request once, keeps quiet while it works, and speaks again only when it has a result, needs a decision or can’t finish. DeepSeek boils the goal down to using the least human attention for the most value.

Yes, but only if your team turns that on. The headlines about an agent that jumps in uninvited are true with a big caveat: unrequested work needs at least one written team rule, and there are three modes.

• Off: it handles direct requests only.
• Observe, the default: it logs the jobs it would have taken, but does nothing.
• Auto: a positive check starts the job.

DeepSeek suggests starting in observe mode, reading through what the agent would have done, and only then switching to auto. Inside the company, it says, one rule let the agent quietly reproduce reported bugs and say nothing until it had both confirmed the problem and prepared an easy fix. Once, it says, the agent had a fix up before the humans discussing that bug in a group chat had finished their conversation.

The short version: Claude Code and Codex are established tools you talk to, while the Super Agent is an experiment in a tool that lives inside your team’s workflow and decides when to help.

• How work starts: with Claude Code or Codex, a developer describes the problem and asks. The Super Agent can also pick up issues, test results and chat messages, if the rules allow.
• Maturity: Claude Code already explores codebases, debugs, runs tests and makes pull requests for many developers. DeepSeek warns its prototype may contain bugs and uneven coding style, and says coding agents did much of the engineering.
• Model choice: the Super Agent uses whatever model your DeepSeek framework profile points to. That defaults to DeepSeek’s own service but can be Anthropic, OpenAI or a model running on your own hardware.
• Where it works: the open version supports github.com and Feishu only. Slack, Teams, GitLab and GitHub Enterprise Server aren’t supported out of the box.

DeepSeek’s own early test hints at the trade-off. In a synthetic run of about 300 messages, roughly 20 of them real work requests, the Super Agent took a median of 12.4 minutes from request to result. A single session allowed to spawn helpers took 20.4 minutes when the model never delegated, and 13.8 minutes when told to delegate. The coding itself was simulated, so this measures coordination rather than code quality, and the numbers are DeepSeek’s. The design also reads far more tokens, which means a bigger model bill.

Getting AI agents to work as a team is still hard. In January, researchers behind the CooperBench study found that coding agents working together had success rates 30% lower, on average, than each agent doing the tasks alone. That study didn’t test DeepSeek’s agent.

Yes, technically. It’s open code on GitHub, and nothing stops a developer anywhere from downloading it. Whether you should comes down to three things.

• Where your code goes. The agent runs on your machine, but the thinking happens wherever your AI model lives. Leave it on DeepSeek’s service and your code, issue threads and the names in them go to DeepSeek, whose privacy policy says it keeps personal data in China. Pointing it at Anthropic, OpenAI or a self-hosted model avoids that.
• How much access it gets. Workers run commands without asking first, with the full permissions of the computer account running them, and can read that account’s home folder. DeepSeek’s own setup guide recommends a dedicated account that holds only the agent’s credentials. Follow that advice.
• Your employer’s rules. US civilian federal agencies were told last month to remove DeepSeek. If you work in government or a regulated industry, check policy first, even if a different model sits behind it.

For most people, nothing changes this week. DeepSeek tells users to review and adapt the code before running it, setup takes many steps, and some ideas from its design essay, such as a watchdog that checks whether each worker is making real progress, aren’t in the public code yet.

The bigger shift is the idea itself. The AI coding race is moving from chatbots that answer to agents that act. Days earlier, at its Universe event, GitHub showed a Copilot app that runs agent jobs across several repositories in parallel. DeepSeek, which has raised about $15 billion ahead of a planned stock market listing, is betting that the winner will be the agent that interrupts you the least.

If you want to try it, keep it boring and safe: a test repository, a separate user account, a short list of people allowed to give it work, observe mode for the first few weeks, and branch protection so nothing merges without a human signing off.