An AI coding assistant can help solve a problem in one conversation. But software work rarely stays in one conversation. A bug report becomes an issue, a proposed fix becomes a pull request, and review comments may arrive long after the first exchange.
DeepSeek's SuperAgent research project explores how an agent might keep track of that longer cycle. The public prototype connects to a GitHub repository, accepts eligible requests, organizes work into separate sessions and can return an outcome to the conversation where the request began.
Exploring AI coding workflows? Visit our Coding Plan page to review available options and their current terms.
SuperAgent is an MIT-licensed research prototype built for DeepSeek Harness. Its repository says it is not an official DeepSeek product or production-ready software. The published code differs from the team's internal implementation, and some ideas in the accompanying research are not implemented in the prototype.
A coding agent with a longer memory for work
The prototype is designed around one GitHub repository, with Feishu as an optional additional source. An authorized teammate can ask the agent to investigate a problem or take on a coding task. The system then creates a work item and starts an agent session in its own working directory.
That organization matters when a team has several requests in motion. A central session handles incoming activity and routing, while separate executors concentrate on individual tasks. The system can associate later information with existing work and report back in the original discussion. A result might include a pull request, which gives a reviewer a concrete change to inspect.
Consider a failing test reported in a GitHub issue. An authorized request could lead the agent to investigate the failure and propose a patch. A reviewer could then examine the changes and provide feedback. This is an illustrative workflow, not a guarantee that every installation will complete each step successfully.
What makes the approach different?
SuperAgent brings task intake, execution and communication into a single workflow. It is designed to distinguish a new request from a follow-up, keep work organized across sessions and surface items that need a person through a web panel.
That does not make it the only coding agent capable of asynchronous work or pull request follow-up. The interest here lies in the way the prototype combines those pieces and in its stated goal of reducing the attention a team spends managing the agent.
Permission comes before initiative
The project describes an agent that acts when an allowed person asks it to, or when a team-written rule permits the action. For work nobody explicitly requested, its documentation recommends starting by recording what the agent would have done before enabling proactive execution.
There are two decisions to evaluate: whether an event should become work, and what the resulting worker is allowed to do. A request to investigate a bug does not automatically justify broad access or every possible external action. Teams still need to set appropriate credentials, permissions and review requirements.
Separate working directories help keep tasks apart, but they are not a security sandbox. Repository content, issue text and logs can contain untrusted instructions. Review remains essential for changes that affect real systems.
Where a team might try it
The architecture could be useful for narrowly defined engineering work: investigating a failing check, preparing a small maintenance patch, or following up on feedback on an existing pull request. Each task needs a clear scope and a way for a person to verify the outcome.
A sensible experiment starts with a low-risk repository, limited access and a specific assignment. Review the changes and tests, but also track the time spent clarifying, correcting and reviewing the output. A quick patch has limited value if it creates more work downstream.
Research code, with practical limits
The maintainers caution that the public code may contain bugs and inconsistent engineering practices. They also state that the released prototype differs from the internal version and does not implement every direction discussed in the research. Features should therefore be checked against the code and documentation rather than assumed from the design article.
Results will depend on the repository, model, permissions, task and review process. The project does not establish a universal productivity improvement, and a separate working directory does not remove the need for human oversight.
How it fits with DeepSeek Harness
SuperAgent is not a new language model. In this setup, a model provides reasoning and generation, DeepSeek Harness provides an agent environment with tools and sessions, and SuperAgent coordinates the flow of repository work. The public repository describes the prototype as a bundle for the harness.
For developers, the project offers a concrete way to study an agent that stays involved after its first answer. Its value in a particular team will depend on whether it makes work easier to review and reduces the effort needed to get a task across the finish line.