Verkkopohjaisten käyttöliittymien ja rajapintojen arviointi suurten kielimallien aikakaudella — Tekoälyagenttien suorituskyvyn vertailuanalyysi ihmiskäyttöliittymissä ja ohjelmistorajapinnoissa
Year: 2026
Type: dissertation
Abstract: The rise of Large Language Models (LLMs) has enabled a new wave of autonomous agents capable of carrying out more complex web tasks than ever before. However, websites are architected for human consumption rather than machine interaction, posing difficulties for agent reliability and efficiency. This thesis evaluates how different web interface design principles affect the operational performance of LLM-based agents on human-oriented and machine-oriented interfaces. The evaluation is performed through a comparative benchmarking strategy, where a custom e-commerce environment exposes the same content and functionality through six different interfaces: a traditional HyperText Markup Language (HTML) site (semantic and non-semantic), a Markdown equivalent of the HTML site, a Representational State Transfer (REST) API, a hypermedia REST API, and a Model Context Protocol (MCP) server. Agent performance is benchmarked across these interfaces in terms of input token consumption, latency, and interaction step count. The results show that the agent-native MCP server achieves the highest efficiency and determinism. For web services, the hypermedia-driven REST API outperforms the schema-driven REST API. In human-facing interfaces, accessible, semantic HTML increases token consumption due to verbosity, but it substantially increases agent interaction reliability compared to non-semantic HTML. Furthermore, serving a lightweight Markdown representation of website content provides a highly token-efficient solution for content-heavy, read-only web tasks. Based on the findings of this thesis, it can be concluded that agents perform more efficiently on machine-oriented interfaces than human-oriented interfaces, and agents greatly benefit from compact content representation and clearly defined available actions.
Source: Aaltodoc (Aalto University)
Author: Koponen, Ruupert
Language: English
Topic: Web Data Mining and Analysis
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204018664