What is Agent Experience (AX)?

Your next website visitor won't have eyes.

It won't admire your hero animation, scan your navigation bar, or feel reassured by your trust badges. It will parse your DOM, traverse your accessibility tree, and attempt to complete a task on behalf of a human who never intends to see your interface at all.

This is already happening. ChatGPT's agent mode, Google's Chrome Auto Browse, Perplexity's Comet, Anthropic's Claude computer use — these aren't research demos anymore. They're shipping products with hundreds of millions of users behind them. And every one of those users is about to delegate tasks to agents that interact with your website the way a screen reader does: structurally, semantically, and without patience for ambiguity.

The question isn't whether AI agents will use your website. It's whether your website will work when they do.

This is the problem that Agent Experience — AX — exists to solve.

Where AX came from

The term was coined by Mathias Biilmann, CEO of Netlify, in a January 2025 blog post. His framing was clear and deliberate: just as UX optimises how humans interact with products, and DX optimises how developers build on platforms, AX optimises how AI agents operate within digital environments.

Biilmann's original focus was on the developer infrastructure layer — clean APIs, machine-readable documentation, context files like llms.txt, and authentication flows designed for non-human users. Netlify was seeing thousands of sites deployed daily by AI coding agents like Bolt.new, and the friction points were obvious: agents that could generate an entire application but couldn't deploy it without human intervention at every authentication step.

The insight was that most software is designed with an implicit assumption that a human is operating it. Remove that assumption and things break in ways nobody anticipated.

Within months, the concept expanded beyond developer tooling. Salesforce's Chief Experience Officer Tiff Zaporteza published a framework treating AI agents as a new category of user within enterprise systems. The community site agentexperience.ax emerged as a hub for principles and emerging standards, with contributors including Kenneth Auchenberg (formerly of Stripe's developer platform) and Eric Simons (CEO of StackBlitz).

By early 2026, the term had moved from a niche concern to appearing in job postings, VC investment theses, and enterprise strategy documents. The progression mirrors what happened with UX itself — a term that started in one discipline became a company-wide priority once the economic consequences became clear.

Two sides of AX

As the concept has matured, two distinct but related meanings have emerged. Understanding the difference matters because they imply different design problems, different teams, and different solutions.

AX as agent design (building agents that work well)

This is the Salesforce and enterprise framing: how do you design AI agents that collaborate effectively with humans? It covers trust patterns, transparency mechanisms, progressive autonomy, handoff protocols between agent and human, and the interaction design of interfaces where both human and AI share control.

If you're building a product that contains an AI agent — a customer service bot, an internal operations assistant, a coding copilot — this is your AX problem. The design challenge is making the agent trustworthy, predictable, and useful.

AX as agent readiness (making products work for agents)

This is the Netlify and infrastructure framing, and it's the one that has broader implications for most companies. It asks: can an external AI agent, one you didn't build and don't control, successfully use your product on behalf of a user?

This framing treats agents as a new audience for your existing product, the same way mobile users were a new audience for desktop websites in 2010. You didn't redesign your business for mobile. You made your existing product work on a new medium.

The same thing is happening with agents. Your customers will send their AI assistants to your website to buy things, book appointments, fill out forms, compare prices, and extract information. If your site works for those agents, you keep the customer. If it doesn't, the agent will route them somewhere else.

IndexLabs operates in this second space. We benchmark how well existing websites work for AI agents — not by checking a list of structural signals, but by sending real agents to complete real tasks and measuring what happens.

What AX looks like in practice

The agentexperience.ax community has outlined a set of principles that are still evolving but form a useful starting framework. At the core: agents are delegates of humans, and the human never stops being the focus. Designing for AX means designing for human needs, delivered through agents as the medium.

Several practical patterns have emerged across the industry:

Discoverability. Can an agent find and understand what your product does? This goes beyond SEO. It includes machine-readable context files (llms.txt, .agent.md), structured data, and clear API descriptions. If an agent can't figure out what actions are available, it can't act.

Navigability. Can an agent move through your product to reach the right place? Vision-based agents see your accessibility tree, not your CSS. If your navigation depends on hover states, JavaScript-rendered menus, or visual cues that don't exist in the DOM, agents get lost.

Content accessibility. Can an agent extract and interpret your content? Client-side rendered applications, content locked behind JavaScript frameworks, and dynamic loading patterns can make your site invisible to agents that don't render a full browser environment.

Task clarity. Can an agent complete a specific task end-to-end? This is the hardest test. It's one thing for an agent to read your product page. It's another for it to add an item to a cart, enter shipping details, and complete a purchase without human intervention.

Why this matters now

Three things are converging that make AX an urgent concern rather than a theoretical one.

Agent traffic is real and growing. Cloudflare's network data shows AI bot traffic increasing rapidly across the web. These aren't just crawlers indexing content — they're agents attempting to take actions. If your analytics can't distinguish agent visits from human visits, you're already flying blind.

The browser is becoming the agent's interface. Google's WebMCP specification (shipping in Chrome 146) lets websites register tools that agents can call directly through the browser via navigator.modelContext. Microsoft co-authored the spec. This isn't a startup experiment — it's infrastructure-level investment in the agent web.

Your competitors are starting to get this right. The companies that optimise for agent interactions first will capture the users who delegate. Every friction point in your agent experience is a handoff to a competitor whose agent experience is smoother.

The parallel to mobile responsiveness is almost exact. In 2010, you could ignore mobile because traffic was small. By 2015, Google was penalising non-responsive sites in search rankings. The companies that moved early had a structural advantage that late movers spent years trying to close.

We're at the 2010 moment for agents.

Where to start

If you're a product or engineering team looking at AX for the first time:

Understand what agents see. Run your site through a tool that renders the accessibility tree and DOM structure. Compare what you see as a human to what an agent sees. The gap is usually surprising.

Test with real agents. Don't rely on structural checklists alone. Have ChatGPT, Claude, and Gemini attempt your most common user task — purchase, signup, booking. Record what breaks.

Fix the structural foundation. Semantic HTML, ARIA labels, server-side rendering, clean heading hierarchy, accessible form labels. These aren't new standards. They're the same accessibility best practices that have existed for years. The difference is that agents make them load-bearing instead of nice-to-have.

Measure and benchmark. AX isn't a one-time fix. Agents update constantly. Your site changes. Treat agent compatibility as a metric you track, the same way you track Core Web Vitals or conversion rate.

The road ahead

AX is in the phase where the problem is clear but the standards are still forming. MCP, WebMCP, llms.txt, and various agent SDKs are all competing to define how agents and websites communicate. Consensus will take time.

What won't change is the fundamental dynamic: humans are delegating more tasks to AI agents, and those agents need to interact with digital products to complete them.

UX didn't replace graphic design. It didn't make visual craft irrelevant. But it changed what "good" meant — from "looks beautiful" to "works for the person using it."

AX does the same thing, one layer deeper. "Good" now means "works for the agent acting on behalf of the person using it."

The websites that understand this first will have a compounding advantage. The ones that don't will wonder why their traffic is declining — without realising that it shifted to a channel they never designed for.

IndexLabs benchmarks how well websites work for AI agents. The AX Index scores 30 leading sites across six categories. See how your site compares →