DeepSeek Harness v0.1: New Open-Source Agent Framework or Just Another Harness?

DeepSeek has released DeepSeek Harness v0.1 in Developer Preview on August 13. It is open-source under MIT license, available as npm package @deepseek-ai/dsh, and runs with a single command. The harness is built on the principle that everything is a plugin, including models, tools, skills, sessions, sandboxes, file systems, the main agent loop, orchestration, and even the interface. It offers four modes: Standard, Code, Minimal, and Creator. The Minimal mode includes only two tools and is designed as a measurement instrument for benchmarking models, minimizing the harness's contribution. DeepSeek's flagship V4-Pro-0813 scored 87.9 on Terminal-Bench 2.1, measured with this harness. The release comes after a week where Meta launched Muse Code and DeepSeek updated V4-Pro, and follows a quarter where xAI, Moonshot, Z.ai, Meta, and DeepSeek all introduced their own harnesses. The article details the anatomy of DeepSeek Harness, its Cordis-based IoC core, its configurable profiles and presets, and its full traceability features. It also discusses the broader industry trend of labs developing their own harnesses for benchmarking, training (RL), and monetization. Security notes include the warning about breaking changes and a PyPI package named deepseek-harness that is unrelated to the official release.