Prime Intellect has open-sourced a self-evolving synthetic environment called general-agent for training AI agents. The core idea is to turn task generation into a two-player game: a synthesizer and a solver alternate, one creating tasks and the other trying to solve them. So far, the system has automatically built a database of 4,504 tasks with over 8,000 unique tools.
The framework starts from simple seed tasks and uses nine strategies — including conditional constraints, noisy instructions, and cross-entity coupling — to split tasks into five difficulty levels (t0 to t4). The synthesizer designs tasks that come with databases, interaction tools, and verification functions; the solver attempts to complete them. Only tasks whose pass rates fall within a specific difficulty range are kept, and the hardest ones become seeds for the next wave of evolution.
In official tests, fine-tuning a 30-billion-parameter model with just 4,400-odd trajectories synthesized in this environment boosted tool-calling accuracy from 18.9% to 52.3% on the BFCL benchmark.
This mechanism frees models from relying on human-annotated static datasets. Through direct game-playing between models, the system can continuously generate training data with controllable difficulty and semantic verification.