Prime Intellect open-source self evolving intelligent agent environment: allowing AI to "fight left and right" and generate over 8000 testing tools According to Beating monitoring, Prime Intellect announced the open source of the general agent training environment, which is a fully synthetic environment that can evolve on its own. The core of this release is to set task generation as a two player game: alternating between synthesizers and solvers. Currently, a large state database containing 4504 tasks and over 8000 unique tools has been automatically constructed. This framework starts with simple seed tasks and divides the tasks into five difficulty levels from t0 to t4 through nine strategies, including conditional constraints, noise instructions, and cross entity coupling. The synthesizer is responsible for designing tasks with databases, interactive tools, and validation functions, while the solver is responsible for attempting to complete them. Only tasks with a pass rate within a specific difficulty range will be retained, and the most difficult level will serve as the seed for the next wave of evolution. Official testing shows that by fine-tuning the 30B parameter model with just over 4400 synthesized trajectories in this environment, the accuracy of tool invocation was improved from 18.9% to 52.3% in the BFCL benchmark test. This mechanism frees the model from relying on manually annotated static datasets. Through direct game between models, the system can continuously generate training corpus with controllable difficulty and semantic verification. [Original link]