DeepSeek Releases Super-Agent Design That Routes Work to Parallel Workers

DeepSeek’s Harness team published a design for a Super-agent on 31 October that aims to keep one agent running across multiple tasks and platforms while limiting the amount of human attention required. The architecture splits work between a Router that handles conversation and dispatch and a set of isolated Workers that execute the actual tasks.

The team identified a practical limit: a person can usefully track only three to five active agent sessions at once. Beyond that point, human attention becomes the bottleneck rather than model capacity. The Super-agent design tries to move routine session management and coordination into the system so that people set goals and review results while the agent handles parallel execution and recovery from interruptions.

In the proposed structure the Router maintains shared context, receives requests from chat platforms or GitHub, and assigns work. Each Worker runs in its own context and directory, carries out the assigned task, and reports status and results back to the Router rather than speaking directly to the user. Additional mechanisms include event-driven triggers, asynchronous operation and a Watchdog process that monitors for failures. The design also supports continuity across messaging and coding platforms so that the same task can move between them without restarting.

DeepSeek released an MIT-licensed research prototype, dsh-superagent, packaged as a bundle for its Harness framework. The prototype is written in TypeScript and is intended for a single GitHub repository. It can optionally listen to Feishu. When an authorised person requests work, it opens a tracked ticket, performs the job in a separate session, and replies in the original conversation, for example with a pull request. Auto-merge is limited to approved pull requests that change fewer than 200 lines and is never enabled under the agent’s own account. The README states that the code is a research prototype, not an official DeepSeek product, and differs from the internal version the team uses.

In DeepSeek’s own synthetic test the Super-agent completed requests at a p95 of 15.7 minutes compared with 29.5 minutes for a single-session baseline. The test used simulated group-chat and GitHub streams of about 300 messages containing roughly 20 real work requests. Coding tools were replaced by time-consuming stubs so the measurement focused on scheduling and coordination rather than code quality. DeepSeek described the results as preliminary.

An internal Super-agent instance is already running inside DeepSeek. It monitors public group chats, attempts to reproduce reported bugs, and can open pull requests even when no one has directly addressed it. The company has not released the full internal system or production performance numbers beyond the synthetic test.

The open prototype and the architecture paper give other teams a concrete reference for building persistent, multi-task agents that do not require a person to manage every session. Whether the Router-Worker split and the measured speed-up hold under real coding workloads with actual code quality metrics remains an open question that the released research code now lets outsiders examine.