The focus on browser automation for AI agents was misplaced. Tools like Moltbot demonstrate the real power lies in an OS-level agent that can interact with all applications, data, and CLIs on a user's machine, effectively bypassing the browser as the primary interface for tasks.

Unlike screen-reading bots, web agents can leverage HTML's declarative nature. Tags like `<button>` explicitly state the purpose of UI elements, allowing agents to understand and interact with pages more reliably and efficiently. This structural property is a key advantage that has yet to be fully realized.

Advanced AI agents like Codex offer a "Computer Use" skill that lets them control your computer's browser and mouse. This is a paradigm shift from traditional automation, which relies on APIs or command-line interfaces. It allows the agent to perform tasks on any application, just as a human would.

Training AI agents to execute multi-step business workflows demands a new data paradigm. Companies create reinforcement learning (RL) environments—mini world models of business processes—where agents learn by attempting tasks, a more advanced method than simple prompt-completion training (SFT/RLHF).

Instead of slowly mimicking human clicks on a website, the "Unbrowse" tool allows an AI agent to learn a site's underlying private APIs. This creates a much faster and more efficient machine-to-machine interaction, effectively building a "Google for agents" that bypasses the human-centric web.

While headless APIs are ideal, many websites and apps actively block headless browsers to prevent scraping. This forces AI agents to interact with the standard graphical user interface to complete tasks, just as a human would, rather than relying on APIs.

By giving agents control over physical or virtual smartphones, they can interact with millions of existing mobile apps via their user interfaces. The Phone Claw concept shows this bypasses the need for specific API integrations, opening a vast, untapped frontier for automation, competitive analysis, and QA testing.

While tech giants may create walled gardens to control AI access (akin to Netflix in streaming), agentic AI has a workaround. Instead of relying on APIs, these agents can take control of a user's browser and interact with websites directly, potentially circumventing platform restrictions.

To overcome the brittleness of UI automation, Amazon's Nova Act uses reinforcement learning in simulated environments called 'web gyms.' These gyms are replicas of typical UIs where the agent self-plays and learns through trial and error. This method, akin to how AI mastered Go, teaches the agent to reason and generalize across changing UIs, a leap over imitation learning.