In defense of agents that click buttons (www.smooth.sh)

A recent discussion highlights the significance of traditional web interfaces in the context of AI agents and machine learning models. While there's a tendency to view APIs as the primary interface for agents, the author argues for the enduring value of established website grammar, which has evolved through decades of user behavior. This "grammar," comprising familiar elements like the logo in the top left and universal navigation patterns, facilitates user interaction and is deeply ingrained due to Jakob's Law. Notably, large language models (LLMs) are trained on this web structure, making them adept at navigating and interpreting these familiar patterns. The article emphasizes that while building AI-driven agents utilizing Machine Code Protocol (MCP) offers flexibility, it risks encountering a steep learning curve for each unique server. For MCP to become as intuitive as existing web patterns, there must be a convergence on shared conventions across servers, a process that may take years. In contrast, agents trained on straightforward HTML can operate efficiently, leveraging pre-existing user-tested structures instead of reinventing them. As a result, utilizing plain HTML might not just serve as an alternative; it could be the most energy-efficient method for agents to interact with the vast web landscape effectively.