Why Claude uses the browser like a drunk intern, and how to fix it (pluno.ai)

Recent developments in AI agent technology have highlighted the inefficiencies of traditional browser agents, likened to "drunk interns" fumbling through user interfaces. Over the past few years, engineers aimed to streamline customer support processes by connecting AI agents to product APIs, but found themselves relying on pixel-based UI interactions. This method is slow and error-prone, as agents mimic human actions like clicking and scrolling instead of utilizing the structured data that underlies these UIs. The reliance on this inefficient interface limits both speed and reliability, as agents struggle to navigate and interact with complex web pages. The proposed solution involves a shift towards leveraging the structured interfaces hidden beneath the visual UI. By developing skills that enable agents to learn from their interactions within a product's backend, like API paths and workflows, a significant enhancement in execution speed and success rates can be achieved. Early benchmarks with a new agent, Pluno, demonstrated a notable performance advantage over traditional browser agents—completing tasks faster and more reliably by drawing on its accumulated knowledge of product-specific APIs. This indicates a paradigm shift for AI agents, advocating for direct interaction with product data rather than limiting their function to UI navigation.