A hands-on test of DeepSeek's Super-agent: it fixed a bug nobody asked it to fix

What the author saw

As Sina Finance reported in a hands-on piece from the author behind the "AI信息Gap" account, DeepSeek has open-sourced a prototype of a "Super-agent", an AI program that takes software chores on its own. The author installed it locally and filed a bug report (a GitHub "issue") without tagging anyone.

The agent replied after 49 seconds. In under 2 minutes a proposed fix (a "pull request", or PR) was waiting for review. It changed 1 line of code and passed CI, the automated tests a project runs on every change. The total bill was 4 fen.

DeepSeek states plainly that this is a research prototype, offered for reference rather than as a product. Keep that in mind for everything below.

Setting it up, step by step

The author's first advice is to give the agent its own GitHub account, because every comment and PR it posts appears under that name. The piece says GitHub's rules allow one extra free "machine account" next to a personal one, for automation only. The author named theirs gap-niuma and invited it into the test project with write access.

The agent runs alongside DeepSeek's own tool, DeepSeek Harness, which works on the web and as a desktop app. The author copied the prototype's code to their computer, opened that folder in Harness, and asked Harness to read the setup file and install everything. Harness handled supporting software such as tmux and the GitHub command-line tool, and asked mostly multiple-choice questions. These covered whether to connect only GitHub or also Feishu group chats, who may assign work, and whether to auto-merge. The author took the recommended answer each time.

One step matters for safety: when Harness asks you to log in, use the agent's account, not your personal one. Setup took about 10 minutes and ended with a dashboard on port 3082. Harness also created an hourly health check that restarts the agent if it crashes, and that check only works while Harness stays open.

From being tagged to acting alone

The first test was a planted bug in a notes app: a single tilde, as in "10~20%", wrongly crossed out the text in between. The author filed an issue tagging the agent. It checks GitHub once a minute, so it answered within about a minute. The fix came with a test and before-and-after screenshots that the agent captured itself using the Chrome browser. It never merges its own work; that step is left to you.

Acting without a tag takes three steps. First, switch on intake of new issues and comments in the dashboard settings. The change only takes effect at the agent's next check, and the author lost 8 minutes waiting before realizing that.

Second, leave it in a default "observe only" mode, where it records what it would do but touches nothing. Here a typo report was marked as workable under a copy-fix rule. A complaint that the page colors were ugly was dismissed, with the reason that a taste opinion without a clear direction needs product and design decisions.

Third, switch the mode to automatic acceptance. After that, an untagged typo report produced the 49-second reply described above.

Teaching it to stay quiet

The built-in rules cover only typos and styling. The "silent reproduce and fix" behavior used in DeepSeek's internal group was not included in this prototype, so the author wrote a rule in plain language. It tells the agent to try to reproduce a reported bug in the background and open a PR only if the bug is reproducible and the fix touches one or two files. If the bug cannot be reproduced, the report is vague, or the fix needs a product decision, it stays silent.

In the test, the agent sent no acknowledgment and spoke up after 6 minutes 55 seconds with one sentence and a PR link. For a deliberately fuzzy report (a save button that "seemed not to work once"), it investigated for 4 minutes 43 seconds, could not reproduce it, and posted nothing. The dashboard marked the ticket as needing no action and recorded why.

Controls, and one catch

By default the agent works on 3 tasks at once, each in its own folder and branch. Every ticket has a stop button, and a pause option holds new tickets without interrupting the current one. It also remembers being told to stop: when the author resubmitted an identical issue, it declined, noting this would redo work the user had cancelled.

The catch is that a direct tag overrides its judgment about size. Asked to add an account system with WeChat QR login and cloud sync, it replied "received" and started working instead of asking questions. After 14 minutes a code branch had been pushed. When the author commented that the feature was off, it stopped within about a dozen seconds and submitted no PR.

On cost, the author cites DeepSeek-V4.1-Flash pricing per million tokens (the chunks of text a model reads and writes): 2 fen for cached input, 1 yuan for uncached input, and 4 yuan for output. Most of this work reads from cache, which keeps it cheap. The prototype watches one GitHub repository at a time and does not support GitLab or Gitee.

What to keep in mind

This is one person's experiment on a small test project, not a benchmark, and the bug was a simple one. The result shows the mechanics rather than proof the agent handles complex code well.

The author also warns that lookalike copies of the repository already exist. Use only the official one, found by searching for dsh-superagent-prototype under the deepseek-ai account.