Can AI agents learn from synthetic worlds grounded in real tools?

Training AI agents to use tools requires diverse, complex interaction data—expensive to annotate by hand. GAIS automates this by grounding synthetic tasks in real Model Context Protocol servers and using structure-guided planning to create adversarial scenarios with logical dependencies. On BFCL, τ²-Bench, and ACEBench, models trained on GAIS data matched or beat officially tuned baselines while using far less data, with performance continuing to improve where other approaches plateau.