Comparison of Agent Interfaces to the Web presented at Göttingen AI Developer Meetup
There is currently a lot of experimentation with different interfaces that LLM-based agents may use to interact with websites. One line of work focuses on agents that interact with HTML pages by clicking on links and filling forms. Second, agents can rely on Retrieval-Augmented Generation (RAG) and interact with a search engine that has crawled and indexed relevant pages. Third, websites can expose site-specific Web APIs that agents invoke via the Model Context Protocol (MCP) , an open protocol that standardizes tool discovery and invocation. Fourth, websites can implement the NLWeb interface, which allows agents to ask natural-language queries that are then answered by the website with JSON data using a well-known vocabulary such as schema.org.
In his talk, Aaron Steiner presented the results of a series of experiments comparing the effectiveness and efficiency of agents using the different interfaces to perform shopping tasks, such as searching for products, comparing product offers from different shops, and finally ordering products that fit the user's needs . The presentation was attended by around 75 persons.