Rethinking the Web for AI Agents

The current web, designed with humans in mind, is struggling to adapt to the rapid rise of AI systems. These systems consume and act on information at a pace far exceeding human capabilities, demanding a new approach to web design. The solution lies in embracing HyperText, reimagining the web as a framework of executable semantics designed for AI agents.

Instead of static pages built with HTML, CSS, and JavaScript, this new web envisions pages as arrays of callable tools. Navigation becomes a process of reasoning, and the concept of “Tool-as-State” (TaS) grants Large Language Models (LLMs) direct access to the runtime environment. This transformative approach opens up a world of possibilities, unlocking orders of magnitude more utility for AI agents.

Imagine a world where a “page” is simply a list of tools, eliminating the need for separate documentation. The LLM only interacts with relevant tools based on the context, minimizing its action space and optimizing efficiency.

Consider an e-commerce example: the homepage offers tools like “search_products” and “select_product.” As the LLM interacts with the website, the tool list evolves. On a product page, the tools “view_reviews” and “checkout_product” become available. This dynamic tool selection simplifies the interaction, providing an intuitive experience for the LLM.

The key to this paradigm shift is the concept of “Tool-as-State.” Traditional software hides state within its memory, but TaS elevates every meaningful state-mutation to a first-class tool. LLMs are then able to observe not only the data but also their own capabilities and how they evolve within the system.

This innovative approach benefits both humans and AI agents. Just as well-designed software gradually reveals options to humans, TaS provides a gradual disclosure of capabilities, ensuring cognitive load remains low. This prevents errors and promotes a more intuitive user experience for both human and AI users.

This evolving manifesto for the web highlights the potential of a new era, one where the internet is designed for the efficient collaboration between humans and AI agents. By embracing HyperText and implementing Tool-as-State, we can build a web that unlocks the full potential of these emerging technologies.

Leave a Comment

Your email address will not be published. Required fields are marked *