Prime Intellect has open-sourced General-Agent — a synthetic environment that generates training data for AI agents without human annotators. Instead of static datasets, it dynamically creates tasks with automatic semantic validation.

The system uses a competitive setup between two models: a Synthesizer constructs tasks with databases and verification functions, while a Solver attempts to complete them. Tasks evolve across five difficulty levels — simple scenarios gradually accumulate additional constraints, cross-references, and complex instructions. The platform retains tasks solved within a target probability threshold, and the hardest cases are used to seed the next generation round. Fine-tuning a 30-billion-parameter model on trajectories collected in General-Agent improved tool-calling accuracy on the BFCL benchmark from 18.9% to 52.3%.

General-Agent: Self-Supervised Data Generation for Agentic Training