Training AI to use the web is seen as a crucial step to expanding its capabilities, potentially making it a more powerful assistant. This trend is evident in recent launches of AI agents integrated into browsers, such as Perplexity's Comet and OpenAI's ChatGPT agent. Microsoft's Edge and Opera are also previewing similar functionalities. Despite high expectations, the author's personal experience with OpenAI's Agent has been underwhelming, often like encountering a novice struggling with the web. The AI's process is transparent, showing each step, but results are frequently slow and inaccurate. Agent often gets bogged down by perceived errors or website malfunctions, and many sites actively block its automated browsing. Privacy concerns also arise, as users must trust OpenAI with login credentials. The AI's potential is limited by its current inability to intuitively understand user needs without explicit instruction. In many cases, Agent is slower than existing web tools and struggles with tasks even suggested by OpenAI. The author highlights APIs as a more efficient method for software integration than AI agents browsing the web. Current AI agents often fail to utilize or properly integrate with existing API features like Connectors. While improvements are expected, the current state of these agents indicates that lofty industry claims of superintelligence outpace actual progress. Mastering web browsing, a seemingly simple task, remains a hurdle for AI superintelligence.