Wang et al. (2026) introduce WebXSkill with three stages: skill extraction from synthetic agent trajectories into parameterized programs paired with natural language guidance, URL-based graph organization for retrieval, and dual deployment in grounded execution or guided instruction modes. Primary source evaluation reports task success rate gains of 9.8 points on WebArena and 12.9 points on WebVoyager over baselines. This directly targets the grounding gap between non-executable textual skills and opaque code-based skills.

Zhou et al. (2023) established in WebArena that LLM agents fail on long-horizon web workflows due to repeated subtasks and error accumulation, patterns also quantified in Yao et al. (2022) ReAct loops lacking intermediate reusable representations. WebXSkill coverage omits explicit comparison to Mind2Web (Deng et al., 2023) hierarchical grounding, which similarly sought reusable actions but without parameterized executability or graph indexing for context-aware reuse. Original extraction from synthetic trajectories assumes high-quality demonstration data, a condition not guaranteed in real-world deployment.

Synthesizing these sources shows skill learning with dual executability and interpretability addresses the core scalability barrier for autonomous web agents, enabling error recovery and adaptation across dynamic internet environments at volume. The URL-graph organization and guided mode together form a hybrid paradigm absent from prior single-mode baselines.