Field Notes

You want a new espresso machine.

Today, that becomes a whole little side quest. You open Amazon. You scroll. You compare. You read reviews from people who may or may not know what good espresso tastes like. You check Reddit. You open a local roaster's site because you feel guilty giving Amazon one more dollar.

Then you get annoyed and close the tabs.

The machine can wait.

Now play the same errand forward a little.

You tell your agent, "Find me a good espresso machine. Manual pull. Under $700. Nothing gigantic. And I don't want some plastic spaceship on my counter."

It already knows enough about you to narrow the field. It knows you care about build quality. It knows you would rather buy once than replace junk twice. It knows you like the ritual of making coffee, which sounds dramatic until you remember people build entire morning routines around this stuff.

So it goes shopping.

Maybe it checks Amazon. Maybe it checks a roaster in Portland. Maybe it checks a kitchen supply store you have never heard of.

You probably do not care.

It brings back three options, explains the tradeoffs, asks for approval, and buys the one that fits.

The machine shows up Thursday.

You never opened a browser.

I would have called that a fun thought experiment a year ago. Then Stripe started putting rails under it.

At Stripe Sessions 2026, they talked about agents as buyers, not as a cute AI demo. Their Agentic Commerce Protocol gives agents a way to complete purchases for people. Link's wallet for agents lets a person approve a spend request before the agent gets a one-time-use card or shared payment token.

Not flashy.

But this is how these shifts usually start. Boring plumbing first. Then everybody wakes up later and acts like the change happened overnight.

The web we know was built for eyeballs and clicks. Product pages. Menus. Forms. Buttons. Hero images. Cookie banners. Checkout flows. The whole thing assumes a human is sitting there, looking at the screen, deciding what to click next.

Quick pause if you are new to this language: an AI agent is software that can use AI to pursue a goal and complete tasks. It does more than answer questions in a chat box. Claude Cowork and Claude Code are versions of this idea. OpenCode is similar to Claude Code, but it can use different models. OpenClaw is the version I care about most because it is not trapped inside one model provider, which ties directly back to Don't Marry Claude.

Agents do not care about most of the visible web.

They want the boring stuff underneath: price, availability, specs, return policy, reviews, shipping time, payment rails, permissions, and whether your system can actually be used by software without a human babysitting every step.

I think people get the first read wrong here.

Your website still matters. Brand still matters. Trust still matters. Humans still browse, compare, and get a feel for who they are buying from.

But the website is not the whole interface anymore.

I think a lot of businesses will miss this because the visible internet is the part we are used to judging. We judge the homepage. The design. The copy. The polish.

An agent judges something else.

Can it read you?

Can it understand what you sell?

Can it compare your offer against someone else's?

Can it complete the job without getting trapped in a weird modal, broken form, missing API, or checkout flow that only works if a human clicks through seven screens?

The small-business version is the part I keep coming back to.

A small roaster with clean data and agent-readable systems may compete in places where the old web would have buried it. A giant with a beautiful website but messy underlying access may lose purchases it used to win by default.

Feels weird. But it tracks.

It also connects directly to what I wrote in Don't Marry Claude. Owning your stack matters because the ground moves fast. But owning your stack only helps if the systems underneath can talk to the next layer of the internet.

I ran into this while building an internal project management tool for my own team.

At first, I was thinking about the normal stuff: tasks, owners, deadlines, dashboards. The usual project-management furniture.

Then I realized the interface was the least interesting part.

If Atlas (my OpenClaw agent) or any of my other agents needed to use the system, every meaningful action had to be callable. Create the task. Assign the person. Change the deadline. Pull the project status. Search the history. Update the note.

That made Monday.com, Asana, and ClickUp feel different to me.

To a person, those tools feel different because the UI is different. One has a board. One has a list. One has a cleaner dashboard. The clicks feel different, so the product feels different.

To an agent, they are much closer to the same business underneath.

Tasks. Owners. Dates. Status. Notes. Comments. Views.

The human sees the interface.

The agent sees the workflow.

Plain English: fewer pages people visit. More systems agents can actually use.

Playbook

The practical version is simple.

Pick the five most common things your customers, clients, or team members do with your business.

Not the fancy stuff. The boring repeat stuff.

Book an appointment. Get a quote. Check inventory. Compare options. Request a refund. Pull a report. Change a reservation. Ask for status. Buy the thing.

Then ask one question:

Could software complete this without a person clicking through screens?

If the answer is no, you found a future bottleneck.

That does not mean you need to rip out your website and become an API company by Friday. Please do not do that. That is how people turn a good idea into a six-month technical swamp.

Start smaller.

Make the important information clean. Prices. Policies. Availability. Product details. Service areas. Next steps.

If you have software, expose the actions that matter. If you have forms, make sure the data goes somewhere structured instead of dying in an inbox. If your team has to manually explain the same thing 30 times a week, the system probably is not readable enough yet.

Think API-first, but do not turn that into a religion.

Keep the website. Keep the brand. Keep the trust.

Then make sure there is something underneath it that software can read and agents can use without a person copying and pasting their way through the whole process.

Stripe is building part of the payment rail. Other companies will build other pieces. The stack is early, messy, and probably going to be annoying for a while. Normal. Every new layer starts ugly before it feels obvious.

But I would not ignore it.

If agents become buyers, browsers, researchers, and operators, I would rather be the business they can understand in ten seconds than the one with the prettiest homepage and a locked front door.

Readable beats flashy in that world.

Orientation

The espresso machine example is simple on purpose.

One person. One agent. One purchase.

Business gets more complicated fast.

A company does not need one agent. It needs workflows that can talk to each other. Sales hands off to onboarding. Onboarding hands off to operations. Operations hands off to finance. Finance catches the weird exception. Somebody escalates when the system gets confused.

That's what I am watching next.

If the agentic internet changes how software interacts with the outside world, workflows change how companies operate on the inside.

Next time, we are getting into Workflows Are the New Primitive.

The unit of business is shifting. It used to be the task. I think it is becoming the workflow.

Comment below and tell me: is your business ready for a customer that never opens a browser?

I read every one.

— Brian