DeepSeek has released the code for dsh-superagent, an AI agent built to take coding work from a software team on one project. An AI agent is AI that can carry out tasks on its own, such as using apps and websites, instead of only answering questions. The company's Harness team posted the code on GitHub on Oct. 31 under the MIT license, a license that lets anyone use, change and share it free of charge. The team says the code grew out of its own day-to-day work, and that coding agents did much of the engineering.
The agent watches one project on GitHub and, if the team chooses, the team's chats on Feishu, a workplace chat app. When someone the team has allowed asks it for work, it logs the request as a task and works on it in its own workspace, apart from everything else. It then answers in the conversation where the request was made, for example with a pull request, a proposed change to the code that people can discuss and review before it is added. A web page puts the tasks waiting on a person first.
Its makers built it around one principle: "Take as little human attention as possible, and give as much value as possible." An agent that can act for a whole team can also waste a whole team's attention, they write. So, the team says, it takes a request only when the request is aimed at it directly and comes from a person on an allowed list, or from anyone in a chat the team opened to requests. Work nobody asked for needs a rule the team has written, and at first the agent only records what it would have done. It is designed to confirm a request once, stay quiet while it works, and write again when a result is ready, when it needs a decision or when it cannot finish.
DeepSeek calls the release a "research prototype for reference only, not an official DeepSeek product." The team says the code may still contain bugs. The public version is not the same as the one the team uses inside the company, and some designs the team described in an article are plans for the future that this code does not include. The agent runs as an add-on to DeepSeek Harness, the company's app for running AI agents, which DeepSeek already offers to anyone in an early public version. Setting it up takes several other programs and access to an AI model from a provider, and DeepSeek advises giving the agent its own GitHub account, because everything it writes appears under that name.
The agent posts comments, opens pull requests and messages people under accounts a team sees every day, and the team warns that whoever runs it is responsible for what it says. The team advises starting with a short list of allowed people and with unrequested work only recorded, and never using the agent for mass pull requests, broadcast messages or busywork. DeepSeek has not said whether the agent will become part of its products. The team says the public code grew out of a version it uses in its own day-to-day work.