ServiceNow details AutoSynthData loop for targeted enterprise-agent post-training

ServiceNow CoreAI has described AutoSynthData, a training-data pipeline designed around the failures of enterprise agents in their actual operating environments. Instead of generating generic synthetic prompts, the system evaluates a target model on diagnostic tasks, uses a stronger teacher to distinguish solvable tasks from dead ends and turns recurring capability gaps into new executable training examples. ServiceNow illustrates the approach with EnterpriseOps Gym and frames each task as a system specification, a user prompt and a verifier.

The task-design rules are unusually explicit. A candidate should be feasible in the available environment, resemble work a user would plausibly request and remain difficult enough to expose a current model weakness. The verifier is expected to agree with the task and environment state, reject incorrect trajectories and still accept different valid solutions rather than hard-code one reference path. AutoSynthData then uses a target phase to create focused examples and a multiply phase to generate variants anchored to accepted samples.

Quality control is part of the pipeline rather than an afterthought. ServiceNow describes solver evaluation, execution of the intended trajectory, positive and negative verifier checks, bounded critique-and-repair cycles and batch-level review for diversity and coverage. That makes the work relevant to Open Model Weights as training-infrastructure intelligence, especially for future checkpoints that explicitly cite such a curriculum or dataset-generation process. It is not itself evidence that any particular model weights, license or training corpus changed.