For twenty years, designing a digital experience meant picturing a person in front of a screen: where they click, what they read, which step they drop off at. AI agents add two new situations. In the first, the person no longer does every step. They ask an agent to do it and review the result. In the second, the user is another agent, reading the docs, calling an API, and deciding whether to keep going or get stuck. Agentic Experience, or AX, is the name for designing both.
What is Agentic Experience (AX)?
Agentic Experience is the design of the user experience when an AI agent sits in the middle: a system that builds a plan and takes actions with some degree of autonomy. It covers two questions. What does a person need when they delegate a task to an agent? And how easy is it for an agent to use a product or service?
The acronym comes from Netlify. In early 2025, CEO Mathias Biilmann published "Introducing AX" and defined Agent Experience as the experience AI agents have when they're the users of a product or platform. Netlify breaks AX into three capabilities: the agent can discover what the product does and how to use it, call it and get predictable results, and find a way forward when something fails.
Salesforce and other product teams use the term agentic experience design for the human side: people working with one or more agents that coordinate tasks among themselves. Anyone building an agent for a business ends up working both sides at once.
Why traditional UX falls short
Interface design rested on a simple deal: the person acts, the software responds. Every screen waits for a click. An agent flips that order. It gets a goal ("reschedule next week's deliveries"), builds a plan, touches three systems, and comes back with a result, often without the person seeing the steps in between.
That moves the problem somewhere else. In a traditional app, you work on the friction of each step. With an agent, the hard part is making sure the person understands what they delegated, what was done on their behalf, and how to undo it if it went wrong. Chat works for asking. For reviewing, steering, or correcting an agent's work, it gets in the way.
Agents are reaching real products fast. In August 2025, Gartner predicted that 40% of enterprise apps would include task-specific AI agents by the end of 2026, up from less than 5% in 2025. Two months earlier, the same firm warned that over 40% of agentic AI projects will be canceled by the end of 2027, citing rising costs, unclear business value, or inadequate risk controls. A lot of agents will reach users soon, and a good share of them will get shut down.
What the data says about trusting an agent
The most concrete study of how people delegate came from Anthropic in February 2026. Measuring AI agent autonomy in practice analyzed millions of interactions with Claude Code, the company's coding agent.
New users (fewer than 50 sessions) let the agent run on full auto-approve in about 20% of sessions. By 750 sessions, that number tops 40%. Experienced users also interrupt more: they stop the agent in about 9% of turns, compared with 5% for new users. And on the most complex tasks, the agent pauses to ask for clarification more than twice as often as people interrupt it.
For designers, the takeaway is that trust builds gradually and oversight changes shape with use. Early on, people approve every action. Later, they let the agent run and step in when something looks off. An interface that only offers "approve everything" or "approve every step" doesn't support that progression. An agent that knows when to ask is part of the design too, as much as any button.
Five UX patterns for designing delegation
Victor Yocco, a UX researcher at ServiceNow, laid out a set of practical UX patterns for agentic AI in Smashing Magazine in February 2026, each paired with a metric. These five cover most cases:
Intent preview. Before any significant action, the agent shows the steps it plans to take, and the person chooses to proceed, edit, or handle it themselves. Yocco suggests tracking how many plans get accepted without edits, with a target above 85%.
Autonomy dial. The person sets how much the agent can do on its own, by task type. For example, the agent can dispute charges under $50 without asking and checks in on everything else.
Explainable rationale. A short, plain-language reason for each action, tied to something the person already decided: "I rebooked your flight because the airline canceled it and you'd pre-approved same-day changes."
Action audit and undo. A timeline of what the agent did, with an undo button and a visible time limit ("Undo available for 15 minutes"). If more than 5% of actions of a given type get reverted, Yocco recommends taking that task away from the agent.
Escalation pathway. When something is ambiguous, the agent asks, offers options, or hands off to a person instead of guessing. Yocco puts a healthy escalation rate between 5% and 15% of tasks.
None of the five is a chat window. They're review screens, settings, and logs, familiar pieces of product design applied to a kind of user that didn't exist before: someone who delegates. The last two also work as governance. We looked at what happens when they're missing in our post on rogue AI agents and agent governance.
When the user is another agent
The other half of AX runs in the opposite direction. More and more tasks reach a product through an agent: ChatGPT comparing vendors, an assistant filling out a form, a coding agent configuring a service. In February 2025, Biilmann told The New Stack that more than 1,000 sites were being created on Netlify directly from ChatGPT every day, and that the integration required specific flows for authentication and handoffs.
A person can work with vague docs. They read between the lines and try things. An agent can stall on that same ambiguity. Netlify says agents need responses with a clear status, a specific error code, and a documented next step. For exposing tools to agents, the standard gaining the most ground is the Model Context Protocol, an open protocol we broke down in what MCP is and why it changes automation.
This applies to companies that don't sell software, too. The agent recommending an accounting firm or an online store reads the website, the pricing page, and the FAQ. If that information is messy or out of date, the agent repeats it as is, or picks someone else. OpenAI just gave businesses a direct way to manage this with verified business profiles in ChatGPT.
What to check if you're designing an AI agent
Write three lists: what the agent does on its own, what needs approval, and what it never does.
Design the review screen before the conversation.
Decide where the action log lives and how each action gets undone. This sits in the agent harness, the layer around the model, not in the model itself.
Define when the agent hands off to a person and what context it passes along, so nobody has to ask the customer the same question twice.
Test your product with an agent as the user. Ask ChatGPT or Claude to complete a task on your site or through your API, and note where it gets stuck.
Track three numbers from day one: plans approved without edits, reverted actions, and escalations.
The first item is the one that gets put off the most, and it shapes everything else. Without those three lists, the autonomy dial has no settings, the audit log doesn't know what to show, and escalation comes down to luck.
At Lumen we design agents that run inside real operations: customer support on WhatsApp, internal knowledge bases that teams query in plain language, ecommerce workflows from catalog to post-sale. In those projects, the experience gets defined up front, alongside the architecture: what data the agent has, what it can resolve on its own, when it hands off to a person, and what the team can see of its work.
An agent that does the job well but that nobody understands ends up switched off. That's why AX belongs in the design work from the first assessment, with the same attention the technical integration gets.
FAQ
What is Agentic Experience (AX)? Agentic Experience is the design of the user experience in products where an AI agent is involved. It covers how a person delegates tasks to an agent, supervises it, and corrects it, as well as how easily an agent can use a product, its website, or its API.
What's the difference between UX and AX? Traditional UX designs the interaction of a person who carries out each step. AX designs delegation: how someone asks an agent for something, how they review what it did, how they undo it, and when a person steps back in. It also treats the agent itself as a user of the product.
Who coined the term AX? Mathias Biilmann, CEO of Netlify, introduced it in early 2025 as Agent Experience, defined as the experience AI agents have as users of a product or platform.
What UX patterns are used in AI agent products? The most cited are an intent preview before acting, an autonomy dial by task type, an explanation for each action, an action log with undo, and an escalation path to a person. Victor Yocco details them in Smashing Magazine (February 2026).
How do I know if my product is ready for AI agents? A simple test is to ask an agent like ChatGPT or Claude to complete a task on your site or through your API and see where it gets stuck. According to Netlify, an agent needs to be able to discover what the product does, call it with predictable results, and get clear errors with a next step.



