The DeepSeek Harness developer preview introduces a modular architecture where everything is a plugin. According to the announcement, an agent is defined as a model plus a harness. The harness is responsible for enabling the agent to understand its environment, use tools, and maintain operation in real-world settings.

The heart of the system is the Cordis kernel, which handles plugin mounting, unmounting, and dependency resolution. All agent capabilities are implemented as plugins: models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and the user interface. Cordis services and events allow these plugins to work together seamlessly.

Developers can compose agents by selecting, swapping, or extending any capability purely through configuration, without altering the DeepSeek Harness source code. To get started, users need to install Node.js and then launch the Web UI with the command npx @deepseek-ai/dsh web. Alternatively, the complete source can be cloned from the repository using git clone https://github.com/deepseek-ai/deepseek-harness.

This approach allows for a high degree of customization and flexibility in building AI agents, staying true to the principle that everything is a plugin.