Abstract
Automated agents can experience inefficiency and brittleness when performing tasks on websites, as some user interface (UI) automation methods can be slow or fragile, and public APIs may be incomplete or unavailable. This disclosure describes a method for decomposing website interactions into executable semantic functions through an offline process. The method can involve tracing a user task to collect interaction data, such as UI events, network traffic, and JavaScript execution details. A large language model can analyze this trace to identify logic and components associated with the task. From this analysis, a self-contained, executable function can be generated. This process can create an API-like interface, allowing agents to perform tasks via direct function invocation, which can improve the speed, reliability, and efficiency of web automation.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Donadei, Emrick; Desineni, Kalyan; Pham, Alan; and Sethuraman, Kaushik, "Decomposition of Website Interactions into Executable Semantic Functions", Technical Disclosure Commons, (September 04, 2025)
https://www.tdcommons.org/dpubs_series/8545