Prime Intellect Open-Sources Its Self-Evolving AI Agent Environment with 8,000+ Tools

AIMPACT update, May 19 (UTC+8): According to monitoring by Beating, Prime Intellect has open-sourced General-Agent, a fully synthetic environment capable of self-evolution. At the core of this release is a two-player game for task generation: an agent called the Synthesizer and another called the Solver alternate in competition. The system has already automatically constructed a large state database containing 4,504 tasks and over 8,000 unique tools. The framework begins with simple seed tasks and divides them into five difficulty levels—t0 to t4—using nine strategies including conditional constraints, noisy instructions, and cross-entity coupling. The Synthesizer designs tasks with databases, interactive tools, and validation functions, while the Solver attempts to complete them. Only tasks whose pass rates fall within specific difficulty ranges are retained, with the hardest level serving as the seed for the next generation of evolution. Official tests show that fine-tuning a 30B-parameter model using just over 4,400 trajectories generated by this environment increased tool-use accuracy on the BFCL benchmark from 18.9% to 52.3%. This mechanism frees models from reliance on manually annotated static datasets; through direct competition between models, the system continuously generates training data with controllable difficulty and semantic validation. (Source: BlockBeats)