ServiceNow CoreAI has introduced AutoSynthData, a pipeline for creating training tasks around the specific weaknesses of an enterprise agent. It compares failures from a target model with successful attempts from a stronger teacher, identifies a capability gap and generates new tasks that exercise the same skill in different situations.
The pipeline is designed for agent environments with real tools, policies and mutable state. Candidate tasks must be feasible with the available actions, resemble plausible user requests and remain difficult for the current target model. A verifier then checks the complete trajectory rather than requiring the agent to reproduce one prescribed sequence of steps.
Verification is central because a syntactically valid task can still be impossible, unrealistic or inconsistent with system rules. ServiceNow says a useful verifier must reject incomplete or policy-violating runs while accepting different valid solutions. As the target model improves, AutoSynthData shifts its curriculum toward capabilities that still fail.
The team demonstrates the method with EnterpriseOps Gym and a released dataset. The approach can reduce dependence on manually authored expert demonstrations, but its output is only as reliable as the environment model, teacher behavior and verifier. Synthetic tasks that encode incorrect assumptions or weak checks could train an agent toward the wrong objective rather than fix it.