AutoSynthData lifts enterprise-agent training 35%

ServiceNow CoreAI says AutoSynthData generated 2,000 synthetic EnterpriseOps Gym samples in about 18 hours and improved Gemma’s mean Pass@1 by 7.2 percentage points, a 35% relative gain. Instead of random prompt expansion, it mines target-model failures, has a stronger teacher propose solutions, and verifies each executable task before training; a second ITSM run raised Pass@1 from 18.77% to 27.18%. For teams measuring agents, the archive’s cost analysis of AWS agent evaluation is the needed companion: synthetic volume only matters when the verifier checks state changes, not text resemblance.