The Race Is On to Redesign Everything for AI Agents

Tina He has always been able to see around corners, and as a writer, designer, and entrepreneur, she’s been actively involved in building the future. I met her when she was running Station Labs, which was building developer infrastructure for Web3 (and for which I did freelance content strategy). She’s now leading a team building developer tools at Base, which acquired Station last year. We’re delighted to feature her work in Thesis. In her piece, she explores a paradigm shift that she’s been witnessing first hand: AI agents are independently selecting vendors, negotiating deals, reading documentation, and writing code. So what happens when agents supplant humans as your primary users and customers? Read on to learn what she sees as the three critical dimensions to focus on that will help you succeed in building for AIs—and a roadmap to four billion-dollar opportunities in this emerging landscape.—Kate Lee 

Was this newsletter forwarded to you? Sign up to get it in your inbox.

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.

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. 

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. 

Beyond user experience—to agent experience

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.

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.

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.

Alongside developer and user experience, a new discipline called agent experience (AX) has emerged. Netlify CEO Mathias Biilmann defines 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.

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.

As AI becomes a primary user alongside humans, the developer tools that win will be the ones that nail the experiences for both people and machines.

The agentic attention economy

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.

Let’s say you’re searching for hiking boots. A human user might be drawn to a blog post titled "Top 10 Hiking Boots" due to its direct headline and appealing images.

An AI agent doesn’t pay attention to any of that. Instead, it analyzes its underlying data by looking for "machine-readable structure," such as specific product names like "Brand X Trailblazer Boot"; key features, such as "waterproof" and "ankle support"; and categorizations like "hiking," "outdoor gear," and "footwear."

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.

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.

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.

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.

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?

A new way to engage

The increasing importance of agentic attention will fundamentally alter user engagement—in fact, it already is: Rep AI's shopping concierge, for example, doesn't just browse stores—it helps visitors navigate a shop  much like a human store assistant would, and then completes transactions. But it only works effectively with businesses that provide structured data.

So traditional engagement funnels no longer apply. Agents don't respond to clever animations or emotional storytelling the way that people do. We have to keep that in mind as we design simultaneously for two audiences: the human end user and the AI intermediary.

Similarly, agentic attention is reshaping advertising in ways that would have seemed bizarre just a few years ago. 

Traditional advertising is designed to appeal to human psychology. In a world rich with agents, your product needs to be positioned for AI comprehension. It's no longer just about brand awareness or emotional appeals—it's about ensuring your digital product is structured in ways that make AI systems confident in recommending it.

Early experimentation with this approach has yielded instructive insights. I modeled how various ways of organizing documentation influence agent comprehension and recommendation confidence. The preliminary data suggest that minor, deliberate adjustments—including adding explicit documents for AI to read, like LLMs.txt—can disproportionately affect an agent's propensity to reference or recommend a product, independent of its intrinsic quality.

As agent mediation becomes more prevalent, the advantage will go to whomever can figure out how to reverse-engineer recommendation algorithms and structure their offerings accordingly. The winners won't be those with the biggest ad budgets—or even necessarily the best products!—but those who best understand how to structure their offerings to align with how AI systems evaluate and recommend.

Four big opportunities for startups

For ambitious founders, the gaps in this nascent ecosystem represent billion-dollar opportunities waiting to be seized. Four territories in particular are ripe for entrepreneurial conquest:

• AI-optimized content systems: The market is desperately awaiting content management systems—like customer relationship management software for people in sales—built from the ground up for dual human-AI consumption. Such a platform should be able to automatically generate semantic metadata. Unlike traditional metadata that might only label content with basic tags, semantic metadata captures deeper relationships and context between data elements so AI can parse meaning more accurately. (Think of it as teaching the AI concepts like "monthly sales" or "total users" instead of having it learn complicated database details.) It also needs to optimize embedding structures (the way in which content is represented numerically so that AI can process it) and interface seamlessly with agent ecosystems. Executing on this effectively can redefine how organizations communicate with one another.
• Composable tools marketplaces: With the rise of MCPs, agents now have the ability to integrate with tools more easily than ever. However, there are still gaps in how agents sort through the tools that are out there and pick the best ones to use. The entrepreneurs who build the definitive marketplace for agent-discoverable APIs won't just capture value; they'll create an app store for agents. Companies like Smithery and Composio have taken some early steps in this direction, but it remains a wide-open space. The first platforms that successfully enable frictionless agent adoption will define this emerging economy.
• Agent analytics platforms: A massive blind spot exists in how agent interactions are measured, creating a perfect opening for startups focused on "AI telemetry." Products that build sophisticated tools for tracking agent behavior, embedding performance, and recommendation confidence will deliver unprecedented visibility into previously opaque processes. Early movers have the opportunity to establish the golden benchmark by which an entire industry measures success.
• AI-to-AI negotiation protocols: Perhaps the most interesting opportunity lies in facilitating agent-to-agent transactions. As autonomous systems increasingly negotiate with one another, they need standardized ways to interact. Companies like Stripe and Coinbase have released tools like AgentKit and Agents SDK that address this need, but there is much more work to be done. 

While building in this space requires navigating novel ethical and operational considerations—like potential "prompt manipulation" tactics and dependency risks (where agents rely on third-party components that may unexpectedly break, change, or be discontinued, disrupting entire workflows)—these challenges only magnify the advantage for founders who solve them first.

Preparing for the AI-first future

Beyond these entrepreneurial opportunities, this paradigm shift demands strategic preparation across organizations of all sizes. Forward-thinking leaders should prioritize these critical areas:

• Better human-AI teamwork: Design workflows where people and AI complement each other instead of competing. Implement clear handoffs, understandable interfaces, and well-defined roles. The goal is simple: Let each do what they do best.
• Prepare for ethics and legal work: Regulations for AI recommendations aren’t yet on the horizon, but they will eventually arrive. Build transparency, explainability, and fairness into your products now. Companies that wait for rules to be forced on them will fall behind. 
• New skills and roles: You need people who understand both your domain and how AI works. New roles such as “agent designer” are around the corner. You might need to train existing staff or hire specialists who bridge these worlds. Don't expect traditional team structures to work unchanged. Christina Cacioppo, the CEO of Vanta, is hiring for a role for using AI tooling to automate and improve all aspects of the go-to-market process, from lead generation to creating sales materials. 

“Agents as users and customers” isn't some far-off possibility—it's happening now. Every week, more developers are using AI to find, evaluate, and implement APIs. More consumers are using AI assistants to make purchasing decisions. More businesses are deploying agents to handle procurement and operations.

The question isn't whether AI agents will become your users—they already are, silent evaluators with algorithmic gazes scanning your interfaces, parsing your documentation, judging your products. If the past decades of the internet were about capturing human attention, the next decade will be about earning agents’ trust. 

To read more essays like this, subscribe to Every, and follow us on X at @every and on LinkedIn.

We build AI tools for readers like you. Automate repeat writing with Spiral. Organize files automatically with Sparkle. Write something great with Lex. Deliver yourself from email with Cora.

We also do AI training, adoption, and innovation for companies. Work with us to bring AI into your organization.

Get paid for sharing Every with your friends. Join our referral program.