Verified Synthetic Web Environments Boost Agent Training.

Web agents hold immense promise for automating intricate digital workflows, but their training is often hampered by the limitations of synthetic environments. These environments, while appearing plausible, frequently contain broken links, inconsistent states, or infeasible tasks, undermining effective agent learning. This research addresses the critical gap between scalable environment generation and the need for trustworthy agent learning. The proposed framework constructs synthetic web environments that are not only executable and auditable but also deeply grounded in a verifiable backend state. Each generated website is represented as a structured scaffold, detailing pages, navigation links, database records, state-change markers, and task constraints. Crucially, the system verifies and repairs structural, semantic, consistency, and feasibility defects *before* policy training begins. During agent interaction, standard UI transitions are executed deterministically, while persistent backend updates are triggered exclusively through validated state-change markers. This design enables the compilation of dense rewards from verified task-progress predicates. Across 500 synthetic environments spanning six domains, this method drastically reduces task-blocking defects and boosts the feasible-task rate from 48.6% to 94.8%. The result is stronger PPO policies and improved transferability to real-world benchmarks like WebArena, WebShop, and MiniWoB++, all without requiring LLM calls during evaluation. This demonstrates that verified synthetic environments can serve as a scalable and reliable foundation for training compact web agents, shifting the focus from mere surface-level plausibility to executable, state-grounded supervision.