返回资讯中心
外部精选
软件工程
#开发者实时流

What Is Agent Experience (Ax)?

Agent Experience is the experience AI agents have when discovering, choosing, and using your technology. Practicing AX means systematically improving that experience. Measurement is how you know whether you did.

Hacker News 最新waldekm2026/11/10 · 7 天前13 分钟阅读

以下正文同步自 Hacker News 最新,版权归原站所有,已转换为易读排版。

Agent Experience is the experience AI agents have when discovering, choosing, and using your technology. Practicing AX means systematically improving that experience. Measurement is how you know whether you did.

Developers now ask AI agents to build with your SDK, API, or CLI, and they judge your technology by what the agent produces. If the agent picks a competing service, calls an API you deprecated, or reports success on a broken build, your technology looks broken, even when your docs are excellent for people. Mathias Biilmann of Netlify coined the term in January 2025 and described it as “the holistic experience AI agents will have as the user of a product or platform.” It applies to any AI agent, from a coding agent in your editor to ChatGPT calling your MCP server.

AX comes down to two questions. The first is propensity: does the agent find your technology and choose it? Give the agent a task that doesn’t name your product, like adding email sending to an app, and see what it reaches for. The second is efficacy: once the agent uses your technology, does it use it correctly? Name your product in the task, and see whether the agent calls your current API or one you deprecated. Your eval tool might use other names, like discovery and selection for propensity, or quality for efficacy.

For both, you score what each run produced and what it cost. Cost is easy to skip and easy to get wrong. For example, Claude Sonnet 5 launched with 33% lower per-token pricing than Sonnet 4.6. Yet when upgrading SharePoint Framework (SPFx) projects, Microsoft’s framework for building Microsoft 365 Copilot, SharePoint and Teams extensions, it cost 3.7x more per run in GitHub Copilot Chat, across 3 scenarios and 15 runs per model (Not all model upgrades are upgrades).

So to recap: propensity tells you whether agents reach your technology. Efficacy tells you what happens when they do. Quality and cost tell you whether the outcome is actually better.

You can’t change the model, and you can only influence the harness, the agent app that runs the model, through its vendor. Everything the agent reads and calls is yours, and the AX stack maps out where that line runs.

Lists of AX best practices, like the principles of AX and best practices for agent-friendly products, are a good place to start. But a change that sounds agent-friendly can still make agents worse with your technology, or more expensive. And the obvious fix isn’t always the best one.

Take SPFx upgrades again. When we ran that task five times in GitHub Copilot Chat with Claude Sonnet 4.6 on Windows, the agent passed only 30 of 80 configuration checks. Telling it to use CLI for Microsoft 365, an open-source tool that upgrades SPFx projects, raised that to 75 of 80. But an instruction like that only works for developers who either know of the CLI or who install the skill that carries it.

The agent’s step-by-step trail showed us something more useful. It fetched the SPFx release notes while it worked, so rather than shipping a skill, we fixed the release notes instead. After that, every run used the CLI without being told, for every developer, with no extra skill installed (Behind SPFx Dev Skills). What we took away from this evaluation is that you must watch how agents use your technology, change what you control, and measure whether it got better. These activities combined are the practice of Agent Experience.

Each result holds for the scenarios and agent profile we measured. Yours may differ, and that’s exactly why you measure.

AX has its own vocabulary for talking about how agents interact with your technology. Here are some key terms that we use regularly.

Developer experience (DX). AX builds on DX, including one of its oldest lessons. Usability conventions help, but you still watch real users to know whether your product works. Agents are users now too, and they’re easier to watch. You can run the same task five times with and without a change, which you’d never ask of a room full of developers.

Generative engine optimization (GEO). GEO is about getting AI search to mention your product. It overlaps with propensity, but AX goes further and asks whether the agent uses your technology correctly once it has picked it.

llms.txt, MCP, and other formats. These are surfaces and formats you can offer agents, which makes them part of AX. Like any best practice, they’re hypotheses until you’ve measured them with your technology and your developers’ tasks.

Here’s a pragmatic approach to start evaluating Agent Experience for your technology.

Already shipped a skill or an MCP server? Start at step 2, and compare runs with and without it.

How to measure AI agent extension effectiveness walks through the method, including the mistakes that produce false signals. Any evaluation beats none, because it gets you out of the “trust me, bro” zone, where the only evidence is a demo that looked fine.

Agent Experience is an evolving field, and there are many resources to help you understand and improve it. Here are a few resources we recommend reading.

正文由 探How 从来源站点 RSS 同步,内容未经改写;遇到排版缺失或需要图片、视频时请以原文为准。

打开原文
去社区讨论 →