DeepSeek Releases Prototype of Super-Agent Coding System That Monitors Group Chats Without Prompting

DeepSeek has released a prototype of a coding agent designed to operate more autonomously than current single-session tools. The DeepSeek Harness team published the design and an open-source research prototype on Saturday. The system, called Super-Agent or dsh-superagent, can monitor group chats among developers, identify issues, and propose solutions without a direct request.

The architecture uses a Router that handles communication and shared context, plus multiple Workers that execute tasks in parallel. The Router assigns work and reports progress. Workers operate independently and return results. The design aims for continuity across messaging and coding platforms, efficiency on concurrent tasks, proactivity in spotting problems, and persistence after interruptions.

DeepSeek tested the prototype with simulated group chat and continuous integration streams. Each run contained about 300 messages, of which roughly 20 were real work requests. In five runs the Super-agent completed tasks at a median of 12.4 minutes and a 95th percentile of 15.7 minutes. A single-session agent baseline took 29.5 minutes at the 95th percentile and often failed to delegate. The test replaced actual coding with a tool that only consumed time, so it measured scheduling and coordination rather than code quality. DeepSeek described the numbers as preliminary.

The prototype is packaged as a TypeScript bundle under an MIT license. It runs as a dedicated profile on DeepSeek Harness and requires Node.js 22.19 or later, git, Python, tmux and the GitHub CLI. It is not an official product; the repository labels it research code for reference. An internal version is already running at DeepSeek. It helps reproduce and fix software problems, responds to code reviews, and improves the Harness system itself. When developers discuss a potential bug in a group chat, the agent can attempt to reproduce and solve it in the background.

The team listed four features: continuity, efficiency, proactivity and persistence. One person can track only three to five active sessions at a time, according to the paper. The Router-Worker split is intended to reduce that bottleneck. Users interact with one agent that handles many tasks and speaks only when a human decision is required.

Bloomberg reported that the prototype aims to rival Anthropic's Claude Code. South China Morning Post and Global Times noted the shift in focus from raw model performance to system reliability and practical application. DeepSeek has not released production benchmarks on real codebases beyond the synthetic timing test.

The release comes as several firms compete to build coding agents that require less human coordination. DeepSeek's version emphasizes open-source research and the ability to act on ambient conversation rather than waiting for a prompt. Whether the Router-Worker design scales beyond the simulated loads, and whether it reduces error rates on actual pull requests, is not yet shown in the public material.

Developers can clone the repository and run the prototype locally under the stated requirements. The system uses whatever model the Harness profile is configured to call. DeepSeek has not announced a timeline for a supported product or for integration into its main offerings. The prototype remains research code, and the company has not published independent evaluations of code correctness or safety properties.