Few weeks ago I decided to let an AI agent handle my weekly grocery run. Full end-to-end. The dream.

It planned meals based on what I had in the fridge. Prepped a shopping list. Matched ingredients to the online store. Added everything to the basket. Even went through to checkout.

And it mostly worked. But the webshop was designed for humans. All that beautiful designer magic, the animations, the hover states, the carefully crafted visual hierarchy that makes shopping “delightful” for people… it just got in the way.

My agent couldn’t find the “Add to Cart” button because it was hidden behind a scroll interaction. It got confused by the carousel of “Recommended for You.” It didn’t know that the oat milk I wanted was on page 3 of the search results, not the first one.

I had to step in and assist. And it made something very clear.

The oat milk problem

The agent bought way too much oat milk. Like, comically too much. Twelve cartons. I have a photo of my fridge that I won’t subject you to.

I ended up going to the shop the next day anyway because I wanted to grab some fresh stuff the agent couldn’t evaluate (produce quality, you know?). I basically turned the whole thing into an experiment and accepted the chaos.

But here’s the thing: even with the oat milk disaster, even with having to step in, even with the webshop friction, I got a tangible glimpse of the future. And it made me want it more.

The trade-off nobody talks about

Most of these shiny demos run on cloud models. Your data, emails, purchase history, that embarrassing thing you asked about at 2 AM. All flowing through someone else’s servers.

People will get more when they lose privacy, and I keep wondering does it have to be like this?

I don’t think it does. Running your own local models is clunky right now, but the trend is clear. The tooling keeps getting better. And at some point, the privacy-preserving path won’t feel like a compromise. It’ll just be the way.

Eventually, everyone’s going to have their own AI assistant.

Maybe it lives in your pocket. Maybe it’s a headset. Maybe it’s an implant, or glasses, or some multimodal thing we haven’t imagined yet. The form factor barely matters. What matters is you’ll have something that knows your preferences, your context, your whole life. Something that can spin up automations that save you actual hours every week. In some cases, generate wealth.

A personal agent that can actually do things, not just chat. Reserve the restaurant you love. Reschedule your dentist when a conflict pops up. Buy the specific oat milk brand you like (hopefully not six cartons this time) without you lifting a finger.

McKinsey published a report on agentic commerce estimating that AI agents could mediate $3 to $5 trillion of global consumer commerce by 2030. Shopping-related searches on generative AI platforms grew 4,700% between 2024 and 2025.

The friction

But we’re not there yet. And the blocker isn’t the technology, it’s the whole ecosystem which is built for humans.

Right now, automations work through three paths:

APIs. Clean, structured, reliable. But most companies don’t want to build them. Maintenance burden, security concerns, pick your favorite excuse.

MCP and friends. Anthropic’s Model Context Protocol, Google’s Agent-to-Agent Protocol, IBM’s Agent Communication Protocol. Basically standardized ways for AI to plug into tools and talk to each other. Better, but still requires companies to opt in and build the connector. Gartner says 40% of enterprise apps will feature task-specific AI agents by 2026, up from less than 5% in 2025.

Just accessing websites. The fallback. Your agent navigates the actual UI, clicks buttons, fills forms. Like a human, but slower and more brittle.

API
Clean, structured, reliable. But nobody wants to build them.
MCP
Standardized protocols. Better, but requires opt-in.
CLI
Command-line interfaces. Scriptable, but limited.
UI Scraping
The fallback. Navigate the actual UI like a human.

Guess which one actually works today for most things? Yeah. The third one. And it’s fine. Until it isn’t.

Designing for two audiences

Designers are going to have to start building for AI agents too, not just humans.

The interfaces that win will be the ones that work for both. Human-friendly and machine-readable. Accessible to people and navigable by agents. That’s a different design problem than we’ve had before.

Just to be super clear on this: I’m not saying we need to make everything ugly and utilitarian. But when my agent can’t find the add-to-cart button because it’s wrapped in three layers of JavaScript animations, something is off.

Agent-to-agent

At Theymes, we’re building player support systems. And I keep thinking about this: at some point, support becomes agent-to-agent.

Not human-to-chatbot. Not human-to-human. Your personal AI talks to the company’s AI. They negotiate. They resolve. They handle the refund or reschedule or whatever it is, and just report back to you. “Sorted. Check your email.”

That’s actually a better experience for everyone? The company gets efficiency. You get your time back. The agents speak the same language, literally and figuratively.

Writing down the above felt very dystopian, as it’s one of those things that makes me think, “Yeah, this could very well be the case in the future.” But also maybe great?

The brand question

Here’s what genuinely keeps me up at night about this stuff.

If my agent is the one “shopping” and I stop visiting websites directly, what happens to brands? Is it just price and availability? 78% of consumers have already used AI for shopping or product research in the past three months. McKinsey warns that 20-50% of site traffic is already at risk.

Do we lose something when the browsing experience disappears? When there’s no store window to walk past, no packaging to pick up and turn over in your hands?

Or, this all will end up giving weekly allowance to our own assistants and then you start getting random stuff from Temu. Six cartons of oat milk was just the beginning.

I don’t know. But I’m pretty sure we’re going to find out soon.