DeepSeek releases dsh-superagent to manage coding tasks in parallel

A shift toward parallel AI work

DeepSeek has released dsh-superagent, a research prototype designed to handle coding tasks more autonomously. Unlike many current AI agents that operate in a single linear session, this tool is built to manage multiple tasks at once. It functions as a proactive agent for a specific GitHub repository, listening for requests from authorized users.

The core philosophy behind this release is to reduce the amount of human attention required to manage AI work. The DeepSeek team argues that human attention is the primary bottleneck in multitasking, as people can typically only follow three to five active sessions at once. By automating the management of these tasks, the agent aims to provide more value with less manual oversight.

The project is currently available as a TypeScript bundle under the MIT license. It is packaged for use with DeepSeek Harness and can run on a user's local machine. To operate, it requires a specific set of software dependencies, including Node.js, Python, and the GitHub CLI.

The router-worker architecture

The system operates using a two-layer architecture consisting of a Router and multiple Workers. The Router acts as a long-running coordinator that handles interactions and dispatches work. When a trigger occurs from GitHub or Feishu, the Router evaluates whether to create a new unit of work called a Ticket.

Each Ticket represents an authorized task with a specific mandate. Once a Ticket is created, it runs in its own isolated Worker session, which has its own directory. This isolation ensures that different tasks do not interfere with one another. The Worker performs the actual coding work and returns the results, often in the form of a pull request, back to the original conversation.

To keep the process efficient, the Router uses a smaller set of tools than the Workers. Furthermore, shell commands are moved to the background after five seconds. This design allows the Workers to continue receiving and processing messages without being blocked by long-running commands.

Performance and simulation results

DeepSeek tested the super-agent using simulated group chats and GitHub CI streams. These tests included about 300 messages, with roughly 20 of those being actual work requests. The team compared the super-agent against a baseline of a single-session agent that calls subagents using a neutral prompt.

The results showed that the super-agent was significantly faster in high-noise environments. In the simulation, the super-agent achieved a p95 task completion time of 15.7 minutes, while the single-session baseline took 29.5 minutes. This represents a nearly 1.9x improvement in speed for completing tasks under heavy message streams.

The team also monitored the load on the Router as message arrivals increased. As the frequency of messages rose from one to 40 per minute, the Router adjusted its batch size accordingly. While per-request latency did increase slightly as the volume grew, the system remained functional under the simulated load.

Safety and future development

Because dsh-superagent runs with full OS-user access, the developers advise users to isolate the environment before use. The system includes an authorization document that defines three different modes for proactive work: off, observe, and auto. By default, the agent stays in observe mode, where it records what it would have done without actually performing the actions.

The current release is a prototype, and the team has not yet measured real-world coding quality. The performance data currently comes from synthetic messages and simulated tools rather than live production code. This means the tool is intended for research and testing of the underlying architecture rather than immediate production deployment.

Looking ahead, the DeepSeek team has proposed several ideas to expand the prototype. These include the addition of an external Watchdog for oversight, the separation of storage and compute resources, and the implementation of an asynchronous ReAct loop. These features would aim to make the agent even more autonomous and robust in complex environments.

How it interacts with users

The dsh-superagent is designed to be as non-intrusive as possible. Once it confirms a request, it stays quiet while it works on the task. It only communicates again when it has a finished result, a specific question, or a failure to report. This behavior is intended to let the user focus on other tasks while the agent works in the background.

Users can monitor the progress through a local web panel that lists only the items requiring human intervention. This helps filter out the noise of the agent's internal processes. By only surfacing what needs a human decision, the tool attempts to fulfill its core rule of taking as little human attention as possible.

The agent also supports subscriptions for ongoing interactions. For example, later comments on a pull request can reach the correct Worker without the user needing to send a fresh mention. This creates a more fluid experience where the agent can handle follow-up tasks autonomously based on the conversation history.