GEAR is an open-source, modular, general-purpose framework for recursive self-improvement (RSI). Positioned in a role similar to Scikit-learn, it enables researchers to combine, swap out, and compare different evolutionary strategies while reusing capabilities for task execution, evaluation, and version management. The framework decomposes parent version selection, modification plan design, task sampling, scoring, and candidate ranking into replaceable modules, with shared infrastructure for recording candidate versions, operating conditions, and evaluation evidence. At the same time, it adopts a meta-learning-based dual-layer self-evolution architecture: the inner layer executes specific tasks to generate experience, while the outer layer aggregates cross-task experience to optimize the reusable Agent operating framework, thereby addressing three common challenges in the RSI field—sampling, candidate evaluation, and diversity management. Experiments show that when the target model weights remain fixed, the Agent operating framework evolved through GEAR significantly improves pass rates and goal completion on marketing, customer service, and coding tasks, and the evolved operating framework can be reused across different inference tiers and models. The framework can serve both as an experimental platform for researchers comparing self-evolution methods and as a component integrated into real business processes to iteratively optimize related workflows through task sets.