Last month, I built an AI agent and set it free to see if it can successfully integrate a tool for me. I’d worked on it and tested it extensively, so I had some idea of what to expect. But still, watching it read through documents for a new tool and then use what it learned to deploy code that actually worked—all on its own—was a heady moment. I thought: We’ve got to start designing everything with agents in mind. Because in addition to millions of humans, your customers will soon be billions of AIs that see the world in a totally different way. (View Highlight)
Autonomous agents are already performing a range of business-critical roles. They provide customer support, select vendors, and negotiate deals. At Base, where I lead a team building tools for developers, I’ve witnessed this firsthand. We built developer tools for humans but found that coding agents were increasingly parsing our documentation—writing code themselves to help with tool integration. Soon, it will be commonplace for agents to work on their own like this, similar to how the one I built did. This demands we reconsider how we build, distribute, and engage with users—be they human or AI. (View Highlight)
First and foremost are developer tools. When you’re designing tools for human developers, you have to think about usability, clarity, and reliability. You must offer documentation that people can read and understand easily, consistent APIs, and supportive communities that help people adopt and integrate your tools quickly. (View Highlight)
As a result, content designed primarily for human appeal—with compelling headlines and attractive visuals but lacking clearly identifiable product information, features, and categories—might become essentially invisible to AI evaluators. (View Highlight)
On the other hand, information that is meticulously organized with clear categories (“ontologies”) and standardized descriptions (“schemas”) becomes highly visible to AI. Imagine a product database for hiking boots where each entry includes structured details about the brand, model, material, and intended use. This can appear as a straightforward list to a human but be organized in a way that an AI can readily understand and utilize it to generate recommendations. (View Highlight)
I experienced a version of this at Base. Our early documentation looked perfect for humans, but AI assistants had trouble telling our different product lines apart, so they often failed at their task. After restructuring our documents for better AI visibility, we saw dramatic improvements in success rate. This reflects broader trends—LLMs are rapidly becoming a source of referral traffic to websites, and some data suggests their output results in better engagement than traffic from traditional search engines. (View Highlight)
This is the essence of what I’m calling “agentic attention”: Since AI agents don’t browse like humans, skimming headlines or pausing on flashy visuals, the determining factors that will make content rise to the top of AI recommendations will be significantly different from what we’re used to on the traditional web. Even tried-and-true SEO tactics will wane in importance as today’s web crawlers are gradually supplanted by agents that will prioritize machine-understandable organization and semantic clarity. (View Highlight)
That creates open and exciting design questions: How do we create experiences that satisfy both human emotional needs and AI structural requirements? How do we maintain beauty and meaning while optimizing for machine interpretability? (View Highlight)
There are three key dimensions we need to focus on, each of which I’ll go into in detail below:
• Designing for agent interpretability
• Optimizing for what I call “agentic attention”
• Creating human-agent collaboration models
The stakes are tremendous, as is the opportunity: Fail to consider your product from the standpoint of an agent and your company risks becoming invisible to these new decision-makers. Do it right, however, and you’ll create brand-new, potentially multibillion-dollar markets for your product. (View Highlight)
Things look different when we know most of our users will be LLMs.
This shift is accelerating with the rise of Model Context Protocol (MCP) servers. MCP lets an LLM-based agent reach out beyond its usual knowledge and use special tools and fresh data from other sources. For example, ChatGPT normally can’t see real-time news, weather, or your calendar. But with MCP, it can check today’s weather through a weather service, or use updated financial data from a financial platform. (View Highlight)
MCP makes this possible by defining clear rules for how the model communicates with external tools and incorporates their responses back into conversations. This standardization is critical for the agent ecosystem, creating a common language for AI-to-service communication. (View Highlight)
Alongside developer and user experience, a new discipline called agent experience (AX) has emerged. Netlify CEO Mathias Biilmanndefines it as “the holistic experience AI agents have as users of a product or platform.” Great AX is when an agent performs a task exactly as you wanted it to, and can perform everything it needs to the first time it’s asked. The process also must be cost-effective, with no human intervention needed. (View Highlight)
Achieving that goal takes careful consideration of several different criteria:
• Onboarding: Agent onboarding involves verifying permissions, providing secure access tokens, and offering structured documentation that AI can interpret.
• Developer kits: When building a software development kit (SDK) for humans, you focus on intuitive APIs, detailed error messages, and comprehensive examples that mirror real-world use cases. Agents, however, need standardized, machine-readable product descriptions, explicit instruction flows, and robust metadata so they can understand and take advantage of your tool’s functionality.
• Interactions and permissions: You need to make sure that when an agent connects to your system, it can prove it’s a good actor, and that anything it does can be audited in case something goes wrong. (View Highlight)
While the internet has traditionally focused on capturing human attention through metrics like search rankings, clicks, and engagement time, recent research from Google DeepMind suggests that recommendation systems may shift toward what the researchers call generative retrieval. In this new paradigm, AI agents are moving beyond simply retrieving items based on past user interactions—someone’s purchase history, for example—and learning to understand content on a deeper level to generate predictions for what will matter to users. This approach allows the AI to identify and recommend relevant items based on their inherent meaning, even for new or infrequent items. (View Highlight)
