In 1993 a cognitive scientist named Don Norman joined Apple and asked for a curious new job title that no one had ever heard before: User Experience Architect.
By the mid-1990s, computers had already been on a path into ordinary life, but the changes to both the industry and culture more broadly had yet to be defined. Up until then computers were built by highly technical people for people with a highly specialized skill set. But the success of the personal computer meant that anyone should be able to easily use one without extensive training.
Norman wanted one person responsible for everything a customer went through with the product, from opening the box to the last click, carefully crafting the experience of users.1
Thirty years later we're headed into a similar uncharted era as technology companies build for a new type of user, one that is machine-powered primarily via AI, called an agent.
Although agents are meant to use applications in very similar ways for users and can simulate them very well, a new world has opened to make applications and information even more useful and retrievable by agents.
The industry is now calling this Agentic Experience or AX.
Agentic experience is the next frontier in growth
In January 2025, Matt Biilmann, who runs the hosting company Netlify, noticed that agents from tools like ChatGPT were deploying websites on behalf of people who had never written code, and observed the platforms getting this new burst of traffic were ones especially friendly to agent use.
He coined the phrase agent experience or AX and rebuilt Netlify's site and onboarding flow so any agent could create and deploy a site with easy account handoff after results were delivered. Sites started arriving at more than a thousand a day through ChatGPT alone. By September 2026 Netlify was signing up 40,000 people a day, two thirds of them new to building for the web.
Agents are particularly skilled at comparing and understanding technical tools, so by the time it reaches a hosting company website, it knows exactly what it wants, and it's acting for somebody who's already decided to buy. Agents are also particularly capable in moving quickly and getting around standard sales flows, essentially moving as much more opinionated buyers as soon as they land than an average user.
Until now most websites would either block agents or serve up a version very similar to what a search engine crawler would see for SEO.
A third option arrived in March, when Stripe and Tempo published the Machine Payments Protocol. The site answers the request with payment instructions instead of a yes or a no, and the agent pays a small amount for the call or the page. Documentation and API access stop being overhead a company gives away and become something it sells, at a price set separately from what signed-in customers pay.
The companies furthest along have converged on the same moves. The first is to drop the signup wall: let an agent finish the job, and let the person claim the account once they see it working. Netlify, Clerk, and Prisma all rebuilt onboarding this way, because someone who has watched the product work converts better than someone staring at a form.
The second is to be usable rather than only readable. Documentation in plain text and an llms.txt file at the root of the site let an agent understand a product in one pass, and a server speaking the Model Context Protocol lets it act on the product instead of writing about it. That difference decides whether a company gets cited in an answer or gets used in the work.
The third is tool design, which now matters as much as API design ever did. A few clear tools beat every endpoint wrapped, each description works better written the way a job gets explained to a new teammate, and every error message should say what to try next. That last piece is Anthropic's, learned building tools for its own agents.2
All of this is the outside of agent experience, everything an agent meets when it arrives at a product as a customer. One year in it has standards in llms.txt and the Model Context Protocol, payment rails for both checkout and per-call access, conferences of its own, and job titles.
Build it, rent it, or run a workforce
Many businesses are experimenting with AI agents right now, and what they are building comes down to three choices about how and where an agent works.
The first choice is to build your own. An insurer writes an agent that reads claims, a support team writes one that answers tickets, and each gets tuned against real cases until it is good at the job. It fits the work exactly, but it also needs continuous maintenance. When it needs a person, it finds them in Slack or Linear.
The second is to rent one. A vendor's agent runs on the vendor's platform, there is nothing to build, and it starts working right after sign-up. These agents may connect to email, calendars, and the other systems for scattered context from multiple third parties, and they are fit well for the job they were built for. The support agent answers tickets and stops there. It will not write the fix or brief the account team, and what it learns stays inside that product.
Both choices leave coordination across your team to be arranged elsewhere and silo agents into workspaces that do not share the context, memory, and intelligence that would make them better over time. Mention the agent in a thread and it picks up the task, does the work, and reports back, and that thread is the whole of its understanding and relationship with the team.
Codespeed is a third pattern, one that defines a complete, autonomous workforce on a single, multiplayer-native platform that contains all of the context and intelligence needed running the end-to-end SDLC or software development lifecycle, from the ticket through to code, review, and the release.
Codespeed agents known as specialists coordinate with each other rather than working in isolated silos, and Lead Dev orchestrates the team, assigning work by fit, keeping it moving, only flagging calls that a person needs to make when needed.
How closely a team wants to stay in that loop with Lead Dev is also up to individual preference. Some teams stay in every step, reading the work as it happens and approving each change. Others hand Lead Dev a long-running objective and come back to a finished result. The same system serves both, and Codespeed's Mission Control and work management tools provide for powerful multiplayer coordination between teams of agents and people.
A true multiplayer agent experience has a profound flywheel effect on quality and thus speed in an AI software engineering setting, also known as a software factory. An agent that can see the decisions already made, knows what the definition of "good" is for its team can move far more efficiently and with more value than one in a silo. It can even proactively surface potential improvements not only to a codebase, but to the broader business, without ever being asked.
Norman's one job title took years to become the department every software company now staffs. Agent experience is one year old, and the companies building both the internal and external halves will decide what the next decade of software looks like.
