DeepSeek Splits Coding Agent into Router and Worker, Cutting P95 Task Completion Time From 29.5 Minutes to 15.7 Minutes

DeepSeek has published a repository named dsh-superagent on GitHub, distributed under the MIT license, with a clear positioning: it is a research prototype.

It is built for one core purpose: split incoming requests from GitHub and Feishu into separate parallel worker sessions for execution. The architecture has two layers: a router responsible for receiving tasks and assigning them, plus a fleet of workers that run assigned tasks independently. This is not a brand new concept, but DeepSeek has tightly integrated it with its in-house models and widely used collaborative tools.

The performance metrics are announced as follows: in DeepSeek's own internal simulations, P95 task completion time fell from 29.5 minutes to 15.7 minutes. P95 refers to the threshold where 95% of all tasks are completed within that time, while the remaining 5% run slower. This metric is more representative than average performance, because it tracks long-tail cases that are usually the most frustrating part of coding tasks.

It is critical to highlight that the phrase "internal simulation" carries huge weight. There is no third-party result replication, no public benchmark configuration, and no disclosure of the specific types and complexity of tasks used in the simulation. The 15.7 minute figure should only be treated as a reference, not a guaranteed performance promise.

So what pain points does this solve? When a single-threaded coding agent runs, if one task gets stuck, all subsequent tasks will queue up behind it. The router-worker approach splits out all parts that can run in parallel -- such as file reading, test execution and dependency check -- to different workers. The tradeoff is that scheduling itself introduces overhead: if tasks are split poorly, parallel execution can end up slower than running them sequentially. That is why the real difficulty of this type of system is not at the worker layer, but in how the router accurately judges which parts can be split and which must run serially.

For developers, the current value of this repository leans toward being a "readable reference implementation". The MIT license allows it to be used for secondary development, but research prototypes typically lack features required for production deployment: permission control, failure retry, cost accounting, and observability. None of these features are included in the current codebase, so we will not overstate its production readiness.

The technical direction behind it is worth attention. The trend of coding agents evolving from "one large LLM working on the entire task alone" to "multiple sessions collaborating" has become increasingly prominent over the past year. By open-sourcing the clearly defined boundary between router and worker modules, DeepSeek at least provides a tangible working sample that developers can modify and build on.

The repository and documentation are available in the original link. Users who want to run it should check dependencies and configuration requirements on their own, and do not expect it to work out of the box. The label of "research prototype" means it proves the feasibility of the technical direction, but does not provide full production-grade support.