DeepSeek releases dsh-superagent to manage complex coding tasks

A new architecture for AI agents

DeepSeek has released a research prototype called dsh-superagent. This tool is designed to handle coding requests by routing them to parallel, isolated worker sessions. It is currently available as a TypeScript bundle under an MIT license.

The project is based on a design document titled "Exploring Super-agent Design: Towards a More Autonomous Agent System." This document outlines how the system moves away from simple linear interactions toward a more autonomous framework for managing work.

The agent is designed to monitor platforms like GitHub and Feishu. When a person with the correct permissions makes a request, the system creates a Ticket. This Ticket acts as an authorized unit of work that includes an immutable Mandate explaining why the task is allowed to proceed.

How the router and workers interact

The system operates using a long-running Router and multiple Worker sessions. The Router handles the initial interaction and dispatches tasks to the Workers. Each Worker runs in its own separate directory and agent session to keep tasks isolated from one another.

DeepSeek uses specific presets for these different roles. The Router uses the dsh-superagent-router preset, which provides a smaller set of tools. In contrast, the Workers use the dsh-superagent-worker preset to perform the actual execution of the tasks.

The system also includes a Subscription mechanism. This allows later comments on a pull request to be sent to the correct Worker without the user needing to mention the agent again. Additionally, shell commands move to the background after five seconds so the Workers can continue receiving messages without interruption.

Testing the super-agent performance

DeepSeek tested the prototype using simulated message streams from group chats and GitHub continuous-integration systems. Each test run included approximately 300 messages, with about 20 of those being specific work requests. The team ran each configuration five times to gather data.

The results showed that the super-agent achieved a median task completion time of 12.4 minutes. In comparison, a single-session baseline using subagents under a neutral prompt took a 95th-percentile time of 29.5 minutes. Notably, the baseline failed to call a subagent in any of its five runs.

As the number of events increased from one to 40 per minute, the average Router batch size grew from 1.0 to 2.8. During this period, the latency for each request increased from 2.5 seconds to 3.9 seconds. These tests used synthetic messages and did not measure the actual quality of the code produced.

Safety and system requirements

The prototype includes several modes for proactive work. Users can turn these off, set them to observe, or set them to auto. The observe mode is the default setting, where the agent records the requests it would have accepted without actually executing any code.

DeepSeek noted that the dsh-superagent runs with full operating-system user access. Because of this high level of access, the company advises that the tool should be isolated before it is used in a production environment. This ensures that the agent's actions are contained.

To run the prototype, users need a specific set of software dependencies. These include Node.js version 22.19 or later (or version 24), Git 2.32 or later, Python 3.8 or later, tmux, and the GitHub CLI version 2.48 or later.

Technical optimizations and results

The developers also looked at how removing direct event pass-through affected the system. They found that this change increased Router decisions by about 30%. It also increased cache-read tokens by about 37%, moving the count from 23.23 million to 31.77 million.

The results suggest that the Router-Worker architecture can handle higher volumes of requests more efficiently than a single-session model. By separating the logic of