Harness²

Harness² from Google Research is an innovative plug-and-play framework designed for developing and evolving autonomous AI agents. It leverages dual recursion—task-level propose-probe-compose refinement and harness-level accumulation of reusable experience—to enable large language model (LLM) agents to improve their execution strategies over time without the need for ground-truth verifiers. This self-improving architecture allows agents to adapt dynamically to open-world tasks, making it ideal for complex, evolving environments. By automating the evolution of prompts, tools, and memory, Harness² enhances agent performance and robustness, pushing the boundaries of autonomous AI development. Its open source nature and emphasis on self-improvement make it particularly appealing to researchers, developers, and AI practitioners seeking cutting-edge, adaptable agent frameworks.