DeepSeek Launches Modular Agent Framework as Open-Source MIT Project

DeepSeek released DeepSeek Harness v0.1, an agent framework designed as a foundation kit rather than a fixed coding assistant. Available under the MIT license, the project embraces a plugin-based architecture where all components—models, tools, storage systems, scheduling, and user interfaces—can be selected or customized without altering the core codebase. The framework is built on Cordis, a specialized meta-framework for spatiotemporal composition. Users can operate in multiple modes ranging from a full-featured standard environment with file editing and web search capabilities to a minimal setup for benchmarking. The platform logs all interactions, including system prompts and context injections, in an append-only session record that supports inspection, forking, and replay. Support extends across multiple model providers, including Anthropic and OpenAI through API keys, alongside AWS Bedrock, Vertex, and custom OpenAI-compatible endpoints. A web interface and Python SDK make the platform accessible to developers.