From UX to AX: Designing for an AI Agent World Learnings about Agent Experience (AX) 14/05/2025 Karin Hendrikse @Netlify APIDAYS ‘25 NYC 

You might already be using AI in your IDE Autocomplete and more complex chat features have been here for a while now. Now we’re seeing AI having the agency to undertake actions for us. IMAGINE 14/05/2025 APIDAYS ‘25 NYC 

TODAY AI Agents 1 How tools, memory and agency turn LLM’s from a fancy autocomplete into Agents. AX Principles AX Principles that can help make our API’s and tools agent friendly. 2 What we did The challenges and solutions we worked on at Netlify and what we learned. 3 Apply AX yourself How you can start diving into the world of AX yourself and make your API’s and tools agent friendly. 4 14/05/2025 APIDAYS ‘25 NYC 

LLM is just a fancy autocomplete LLM’s are trained on6 4 cod@ 4 documentatioX 4 bookC 4 websites To predict what comes next in a sequence of words. And… that’s it. But it appears to be pretty powerful! WHAT IS AN AI AGENT? 14/05/2025 APIDAYS ‘25 NYC 

When AI becomes an Agent Give AIE @ ToolB @ Memor0 @ The ability to make decisionB @ Access to API’B @ Access to the file system Suggesting code Running & deploying code WHAT IS AN AI AGENT? 14/05/2025 APIDAYS ‘25 NYC 

Expand your UX and DX with AX Agents are becoming more ‘common’ delegates for the users of our applications. Are our apps ready for them? We already know& ! User Experience (UX$ ! Developer Experience (DX)B ! Agent Experience (AX) A better AX also means a better UX and DX WHAT IS AN AI AGENT? 14/05/2025 APIDAYS ‘25 NYC 

AX: how easy it is for an AI Agent to use your app or API WHAT IS AX? 14/05/2025 APIDAYS ‘25 NYC 

WHY AX MATTERS AX is embracing agents using your applications Adoption Allow agents to use your services so they build preference in recommending it to their users. User productivity and satisfaction If your user wants to use an agent, make sure they can! Build preference for your target audience. Nice extra’sŠ € Better API’s = happy devx € Better Docs = happy devs and power userx € Better security = happy userx € Streamlined workflows = happy users 14/05/2025 APIDAYS ‘25 NYC 

Core AX principles Xoxo - the AX community 14/05/2025 APIDAYS ‘25 NYC 

Human Centricity What agents do needs to align with delivering value to the human If you enable agent access, do it in ways that preserve user trust and intentf 9 Delegated actions taken by the agent are traceable back to the user’s intentf 9 Make sure the user can see, review, or revoke what the agent is doing. We only need to apply AX to the things a human wants to do. AX PRINCIPLES 14/05/2025 APIDAYS ‘25 NYC 

Agent Accessibility Remove technical barriers that prevents agents from using your service If your user can do it, an agent should be able to do it as well This could be through9 8 API’S 8 CLa 8 A well structured website that can be parsed Only use CAPTCHA when you have to... Requiring a human step where it’s not needed is called out as an anti-pattern in AX AX PRINCIPLES 14/05/2025 APIDAYS ‘25 NYC 

Contextual Alignment Teach the agent what they need to know Eliminate the guesswork and prevent misunderstandings. Machine-readable documentatioa 3 llms.txt fileV 3 Context fileV 3 Model Context Protocol (MCPY 3 OpenAPI specs AX PRINCIPLES 14/05/2025 APIDAYS ‘25 NYC 

Agent Interactivity Patterns How can an agent safely handle actions for the user? Establish standard ways for agents to perform actions, especially sensitive. Make sure the agent gets clear scoped API’s, document expected behaviour and keep users in control. Think ofz y Using a special endpoint that triggers confirmation to the userl y Content creation for the user that follows guidelines for tone or provide citations. Maybe using existing tech like OAuth. AX PRINCIPLES 14/05/2025 APIDAYS ‘25 NYC 

Differentiate Agent Interactions Make sure to track what actions were done by agents vs users Apply observability and accountability! Like a user-agent string 
 ‘agent-’ in the User-Agent header. Why? Debugging, analytics, security and detecting patterns for agents. AX PRINCIPLES 14/05/2025 APIDAYS ‘25 NYC 

Essentially: make your platform as easy, transparent, and safe for an agent to use as it is for a human. 14/05/2025 AX PRINCIPLES APIDAYS ‘25 NYC 

Solution Created a workflow where agents can deploy sites to Netlify that the user could then claim themselves. Agents create >60k sites per week Through agents like Windsurf, Bolt.new and Same.new Problem 1 How can we update our API’s so agents can interact with them? PROBLEM SOLVING AT NETLIFY 14/05/2025 APIDAYS ‘25 NYC 

Solution f Add llms.txt to our docsG f Provide netlify context files We even integrated this into our Netlify CLI: one command netlify recipes ai-context will drop the latest context file into your repo. Problem 2 How can an agent know what information is important? Agents might only read part of your website if they scrape it. PROBLEM SOLVING AT NETLIFY 14/05/2025 APIDAYS ‘25 NYC 

Solution Context files Adopt broader efforts like the Model Context Protocol (MCP) Netlify’s own MCP server is coming soon! Fun extra: you can now build your own MCP server on Netlify. Problem 3 Ensure agents know the latest and unlearn the wrong information? PROBLEM SOLVING AT NETLIFY 14/05/2025 APIDAYS ‘25 NYC 

PROBLEM SOLVING ELSEWHERE We’re not alone though... Clerk: redesigning flows so agents can log in on behalf of users. Neon: hiring AI engineers to “dogfood” their product as if they were agents, to make sure it’s agent-friendly. Resend: reducing the amount of friction for agents to send emails. 14/05/2025 APIDAYS ‘25 NYC 

So... why should you care? Your next “user” might be using an agent When agents become commonplace - “Find me a red jacket under $100 and buy it” - might be something your e- commerce site will want to support. Product improvement It’s like an extreme usability test: if an agent can use it, certainly a human can! Ethical and controlled AI deployment One interesting implication of AX – if done right – is that it can actually make AI interaction more transparent and safer. Take ownership of this new change in tech As an engineer, adding AX design to your skillset is going to be valuable. 14/05/2025 APIDAYS ‘25 NYC 

Evaluate your current project’s AX Take something you’ve built – maybe a web app or an API – and imagine an agent trying to use it. Reveal low-hanging fruitv u Maybe just adding proper alt text or data attributes could help an AI extract info from your page.‘ u “we have an API; an agent could theoretically call it. Let’s add an OpenAPI spec.” EXPLORE AX YOURSELF 14/05/2025 APIDAYS ‘25 NYC 

Use AX principles and guidelines to guide your learnings Check out the “Principles of AX” guide on the new website: https://agentexperience.ax/ And browse the articles by the AX community. EXPLORE AX YOURSELF 14/05/2025 APIDAYS ‘25 NYC 

Start small: make something AI-ready Try creating a simple llms.txt for your project (basically a one-page summary a chatbot could read). Or, if you have an API, consider generating an OpenAPI spec if you haven’t – tools like GPT-4 can read those specs when you provide them. If you maintain a library or SDK, maybe add a few comments in the README specifically guiding Copilot on usage (yes, people do that!). EXPLORE AX YOURSELF 14/05/2025 APIDAYS ‘25 NYC 

Experiment with an AI using your tool Actually run a test: use a tool like ChatGPT or GitHub Copilot Chat and see if you can get it to accomplish something with your product. See where it fails or gets confused. This experiential approach will highlight AX issues quickly. It’s like UX testing but the tester is an AI following your instructions. EXPLORE AX YOURSELF 14/05/2025 APIDAYS ‘25 NYC 

Think about ethical design As you explore AX, consider the ethical dimension: what shouldn’t an agent do on your platform? Designing for AX isn’t just opening all doors; it’s putting the right doors in place – some wide open, some requiring a key – and clearly labeling them EXPLORE AX YOURSELF 14/05/2025 APIDAYS ‘25 NYC 

Thank you! Have fun exploring AI Agents! 14/05/2025 APIDAYS ‘25 NYC 

RESOURCES $ https://agentexperience.ax/ - new community site where people are gathering knowledge€ $ https://clerk.com/changelog/2025-03-7-clerk- agent-toolkit - Clerk introduces their AI Agent toolkiž $ https://www.anthropic.com/news/model-context- protocol - Anthropic introducing the model context protocol€ $ https://developers.netlify.com/guides/write-mcps- on-netlify/ - How to build an MCP on NetlifQ $ https://harper.blog/2025/02/16/my-llm-codegen- workflow-atm/ - A delightful blog on codegen workflows (less AX, but still useful imoÇ $ https://www.layered.dev/mcp-the-ultimate-api- consumer-not-the-api-killer - Good read about MCP and API’s RESOURCES 14/05/2025 APIDAYS ‘25 NYC @khenhey.bsky.social $ Book: Co-intelligence, living and working with AI by Ethan MollicQ $ https://cookbook.openai.com/examples/ function_calling_with_an_openapi_spec - function calling using an OpenAPI spec