DeepSeek Harness Hands-On: I Forgive Its Price Hike

DeepSeek finally released and open-sourced DeepSeek Harness (DSH), an agent harness that the author had been beta-testing for nearly half a month. The harness is built on a pluggable architecture called Cordis, which allows users to customize everything from models and tools to strategies and storage, making it highly modular and DIY-friendly. With over 100 built-in plugins, DSH is seen as the 'Android of the Agent era' compared to Codex's more closed 'iPhone' philosophy. In hands-on tests, DSH-led models showed significantly better long-horizon task performance, such as generating a 3D model of Pig Bajie without complex prompts, taking 20 minutes with DSH versus 6 minutes with Codex, though Codex produced a less realistic result. The harness includes features like a trajectory view for debugging, skills for code review and documentation, and a token usage dashboard that showed high cache hit rates around 99%. DSH supports custom presets and allows users to create plugins, with a Plugin Store promised. The author concludes that DeepSeek models are best paired with their own harness, and the open ecosystem could enable community-driven evolution of AI agents.