Recent advancements in large language models (LLMs) and small language models (SLMs) have opened up new possibilities for automating everyday tasks on the web. However, it is still not clear how these agents should best interact with websites, given the different approaches, such as reading the website’s code and structure directly or using structured tools. Especially for SLMs, given their reduced memory and reasoning power compared to larger models, it is unclear how well these different methods work.

This thesis will study how well SLMs can complete interaction-based tasks in the web, using different established methods. Using one or more test websites and a prototypical implementation, a comparison of tool-based interaction with other interaction methods for SLM-based web agents will be conducted. Appropriate qualitative or quantitative evaluations will provide insights into the benefits and drawbacks of each approach, as well as the factors that impact them.

The goal of this thesis is to compare the different ways an SLM-based web agent can interact with a website. This comprises the analysis of the state of the art of web agents, including web automation and web parsing approaches, as well as other relevant literature. Prototypes for one domain should be implemented to test the different methods. A suitable evaluation has to be conducted that considers both quantitative and qualitative aspects of the approaches, as well as their compliance with requirements extracted through the literature research.