DeepSeek releases open-source coding agent that fixes bugs without being asked

What DeepSeek released

As technews.tw reported, the Harness team inside the Chinese AI company DeepSeek published a prototype called Super-Agent on its WeChat official account on October 31. It is a coding agent, meaning an AI program that reads, writes and changes code, and it is open source.

The headline claim is that it can notice a problem in a developer group conversation and fix it without anyone telling it to. The outlet describes it as aimed squarely at Anthropic's Claude Code and at coding tools from US companies such as OpenAI.

Why autonomy is the point

DeepSeek says the main difference from other coding agents is independence. Most work inside a single session: a developer gives an instruction, the agent carries it out, then stops and waits.

DeepSeek's argument is that when several tasks run at once, the limit is no longer the model's ability but the human's attention. Its stated goal is to demand as little of that attention as possible.

How it is built

Super-Agent has a Router layer that holds the overall context and handles communication, plus several Worker sessions that run different tasks in parallel. This lets the agent follow a task across different channels.

DeepSeek says an internal version can trace a problem raised in instant messaging all the way to GitHub, a popular code-hosting site, while also watching public developer discussions. If someone reports an error in a group, the agent can recognise it unprompted, try to reproduce it, and submit a pull request once it has a fix.

The Harness team was formed in May, and a developer preview of DeepSeek Harness was released under the MIT license in August. A harness is the software wrapped around a large language model that coordinates tasks and connects outside tools. Claude Code uses similar technology.

Timing and open questions

The report notes the timing is sensitive. Anthropic has filed for a stock market listing, and DeepSeek has raised a large new funding round that also paves the way for going public. The competition between US and Chinese AI firms is moving from chatbots toward autonomous agents.

The report is not DeepSeek's documentation. It does not say which model powers Super-Agent, what license covers it, or what hardware it needs. Open-source code is also not the same as freely available model weights, so check the license before using it for client work.