Many AI-powered web agents currently rely on raw HTML or visual page representations to interact with web applications. A new paradigm is emerging in which web applications expose their functionality directly as structured, schema-driven tool sets, such as Anthropic’s Model Context Protocol (MCP) and the W3C WebMCP proposal. This more constrained setting raises the question of whether Small Language Models (SLMs) can serve as a dependable reasoning core, enabling a local, low-cost, and privacy-preserving deployment. However, it is unclear how characteristics of the exposed toolset, such as size, schema complexity, or functional overlap affect tool selection of an SLM, and whether optimization strategies applied at the interface between model and application can mitigate it.

This thesis aims to address these open questions by systematically investigating how tool set characteristics affect the reliability of SLM-based, schema-driven web agents. To this end, a controlled experimental system is to be designed and implemented, consisting of an agentic system in which an SLM acts as the reasoning core, interacting with a schema-driven web application that exposes a scalable tool set. Meaningful tool set optimization strategies are researched, integrated into, and evaluated within this system to mitigate performance limitations identified for varying tool set conditions.

The objective of this thesis is the design and implementation of a controlled experimental system for studying SLM-based, schema-driven web agents, followed by a suitable empirical evaluation of how tool set characteristics affect their reliability. This consists of an analysis of the state of the art on web agents, schema-driven tool interfaces, tool selection optimizations, and the use of SLMs in agentic settings. The thesis further includes a prototypical implementation and an empirical evaluation of the impact of tool set characteristics and optimization strategies.