Official agentic environment from DeepSeek built on the Cordis microkernel.
DeepSeek Harness (dsh)
https://github.com/deepseek-ai/deepseek-harness
https://www.deepseek.com/harness/en/
Announced on August 13 alongside the DeepSeek-V4-Pro-0813 release, the official open-source harness from DeepSeek — dsh — is powered by the Cordis microkernel. It was originally created by developer Shigma as the core engine for the Koishi chatbot framework. Later, it was spun off into an independent project (Cordiverse), with researchers from Peking University and DeepSeek formalizing its mathematical model in a paper on "spatiotemporal composability." The paper describes the ability to dynamically load, update, and unload plugins at runtime without restarting the process.
The architecture follows an "everything is a plugin" philosophy: the agent loop, tools, sandbox, and UI are all implemented as modular components with reversible effects (clean side-effect rollback upon unloading) and a shared context. It is MIT-licensed and currently in early developer preview (with potential breaking changes to the API). The repository includes tutorials for building custom extensions, and the community is already publishing additional plugins under the dsh-plugin tag.
https://www.youtube.com/watch?v=RWp5cejTApU
In a hands-on review, the author demonstrates the practical workflow inside the dsh Web UI. Key takeaways from the video:
- Built-in Marketplace (Harness Market) — installing extensions, extra tools, and third-party model providers directly from the interface.
- Operating Modes (Code, Minimal, Creator) — notably the Creator mode for configuring custom agent presets and inspecting runtime state.
- Workspace Management & Live Control — organizing projects by folders, along with the ability to pause execution and adjust goals on the fly without restarting the session.
Discussion
https://news.ycombinator.com/item?id=49285244
On Hacker News, the project is discussed primarily as an example of agentic harness engineering. The main takeaway from developers is that a well-designed environment (feedback loops, context control, and tool invocations) currently yields just as much impact as upgrading the model itself.
It is also compared with Reasonix (a compact Go-based agent optimized for DeepSeek, MIT): Reasonix aims for maximum cost efficiency through stable prefix caching for long-running sessions (often less than $1–2 per task), whereas dsh serves as a more versatile, extensible, and official framework.
#deepseek #harness #agenticcoding #cordis