DeepSeek AI has released DeepSeek Harness (dsh), an open-source agent framework built on a modular plugin architecture. The system operates on top of Cordis and allows users to construct agents by layering plugins, such as session storage, sandboxing, and context compression, using a patch-based configuration system.
The framework supports four agent execution modes: background sub-agents, context inheritance for sequential tasks, JavaScript-based workflow orchestration, and the 'ralph' mode, which uses the file system as long-term memory. Configuration is managed via profiles that can be updated on the fly without restarting the process.
Security is handled through a workspace sandbox that restricts file operations, while session history is saved to files to allow for resumption and token usage tracking. The project is currently in developer preview and is distributed under the MIT license, allowing users to swap models and tools as needed.
Users can start with a basic setup via the web interface at http://127.0.0.1:3080 or headless mode for CI environments. The framework enables the creation of custom pipelines, such as the provided four-step text processing example involving brief generation, prompt polishing, diff visualization, and stylistic adjustment.
