Google and DeepSeek release open-source AI agent frameworks

Google Research has open-sourced Harness², a framework designed to enable AI agents to improve their performance without the need for retraining underlying model weights. Instead of fine-tuning, the system automates the optimization of the model's environment, including system prompts, guardrails, memory, skills, tool guidance, scripts, subagents, and plugins. This approach aims to allow agents to adapt to unanticipated tasks in open-world scenarios by modifying the orchestration around the model.

Similarly, DeepSeek has released DeepSeek Harness (dsh), an open-source agent harness that wraps language models to enable multi-step task execution, such as file reading, command execution, and web searching. Unlike traditional harnesses with a fixed core, dsh utilizes a plugin-based architecture called Cordis. In this system, every component—including the model adapter, tool registry, session log, and the agent loop itself—is a plugin that can be swapped via configuration.