Prime Intellect has released General-Agent, an open-source, self-evolving AI environment designed to enhance task generation and model training. The system features a two-player game involving a Synthesizer and a Solver, which compete to create and solve tasks. This environment has generated a database of 4,504 tasks and over 8,000 unique tools, categorized into five difficulty levels from t0 to t4. The framework uses strategies like conditional constraints and noisy instructions to evolve tasks. The innovative setup allows for the automatic generation of training data, eliminating the need for manually annotated datasets. Tests indicate that fine-tuning a 30B-parameter model with trajectories from this environment improved tool-use accuracy on the BFCL benchmark from 18.9% to 52.3%. This advancement highlights the potential for AI models to evolve through direct competition, continuously generating data with adjustable difficulty and semantic validation.