Asymmetry

TL;DR: Agents have gotten good enough at almost everything, except buying things. They still shop like humans: screenshots, search results and checkout pages built for people. Smarter models don't fix that. Better data does. Give agents live, two-way access to every good and service, and they'll make better buying decisions than any human, in under a second. This unlocks a market where the best products win on merit, great merchants finally get found, decisions are quicker, and everyone gets to live and present themselves exactly the way they want.

Agents are bad at buying things

I work at a startup which sells to companies building agents. I am incentivized to try new AI things. I am also incentivized to over-trust them, so I can catch them failing, so we can productize the fix.

I have been using all the personal agents since day 0, and I believe they have gotten good enough at most of the things they do. Commerce isn't one of them.

Taste, or quality of decisions

Instinct helped me find a doctor and booked an appointment for me. It ended up being neither good nor insured. Not all service appointments are going to be bad, but it still can't do better than me, in probably the same amount of active time I spent texting it.

I used Grok, Instinct and Muse to shop for new sheets for my room. None of them found something of my taste, or something of higher quality than the first Google link. When I tried to check out, it got blocked and handed it over to me.

We have run tons of examples, and the current success rate and confidence aren't ready for the market.

Beyond satisfaction, there are a few other things that kill the experience for me.

Speed

It takes up to tens of minutes to surface the first result. With the follow-up conversation, buying a single thing can take 10–40+ minutes. That's more than the time an average human spends to buy simple things.

It misses the important parts

There's a lot of metadata on shopping websites that factors into the decision: delivery time, shipping cost, discounts, return policies, reviews, stock, seller trust, warranty, sizing and fit, and price history. Some perspective on these helps overall, and I would say beats the in-person experience.

Friction

It's moderately stressful to onboard agents onto shopping infrastructure. Half of them are banned, and the other half are dicey in the best case. It's slightly easier to let them control the search, and outsource checkout to a human.

Here's what we believe is the bottleneck.

Agents are shopping on human infrastructure

Computer use is just a proxy for a better solution

A computer use agent shops like a human. It takes a screenshot, figures out what's on the page, decides where to click, clicks, waits for the page to load, and takes another screenshot. This sometimes happens tens or hundreds of times for each purchase.

It's one of the major things that make agents slow. Each round takes a few seconds, and the process is mostly sequential across hundreds of these human actions. That's why it takes minutes to surface the first result.

It constrains agents. The agent only sees what fits on a screen: the same 10–20 listings a human would see, in the order the website wants to show them. Pop-ups, captchas and login walls stop it cold, and it only moves forward after getting past them.

It's technically inefficient. Every screenshot costs 1,000–1,800 tokens, on top of ~4,500 tokens of tool overhead on every single request, and the agent re-reads the whole page every time something changes. It's compute intensive in a way that just doesn't scale. And it cold starts every time: it doesn't remember it checked the same page yesterday, so it starts from zero.

This benchmark, OSWorld-Human, put a clock on these agents. Stuff that takes a human a few minutes takes an agent tens of minutes. And it's not the clicking that's slow, it's the thinking in between: planning, second-guessing, checking its own work. It gets worse the longer it goes: later steps take up to 3× longer than the first ones. Across 16 agents, even the best took 2.7–4.3× more steps than a human would.

Shopping is worse. On DeepShop, a benchmark of real shopping tasks, the agents Microsoft tested got it right 11.6–32% of the time. And that's before the websites that block them for being bots.

Why web search APIs won't work

The obvious fix is to skip the browser and hand agents search APIs like Exa. It's faster, but it's built for reading the web, not buying from it.

Stale data. Prices, stock and delivery times change by the hour, and a search index might only catch up later. An agent quoting yesterday's price is worse than no agent at all.

It's a one-way street. Search lets an agent read about a product, but action is limited to a hand-off to a browser, or back to a human.

No clear signal of reliability. Search just relays the internet, but a purchase depends on it being right. If Amazon lists a wrong price, it's like blaming Google: no one takes the blame. So the agent double-checks everything itself, and that slows everyone down.

What if we gave smart models better data?

Agents could find the best product in under a second

This alone unlocks:

• Higher conversion for merchants: the gap between wanting something and having it is where sales die. Fulfill the need instantly and more of them convert.
• Time saved for consumers: the 10–40 minutes agents and humans spend going back and forth on one product go back to us. Humans eventually save time, as it would become faster than current online commerce.
• More consumption: when acquiring a product or a service is effortless, you buy the things you'd otherwise keep putting off. (Maybe in a good way?)

Agents could see every listing (a million+) for a given offering

Today an agent sees the same 10–20 listings I do. What if it saw all of them, and only the perfect listing got chosen?

• Higher satisfaction for consumers: you get the one that's actually right, not the one that ranked first.
• Higher confidence in agents: once it picks right a few times, you stop double-checking it.
• A long tail of production: when niche needs finally surface, someone gets to build for them.

Agents could understand taste

• Happiness: users would be happier, and buying would feel joyful again: an experience, not a chore.
• Presentation: when finding the right thing takes a second instead of a Sunday, everyone gets to present themselves exactly the way they want. Society gets a lot more interesting when everyone expresses themselves better.

Agents could make the best decisions possible

• Timing: data that tells them exactly how a price is moving and when to buy, programmatically.
• Real satisfaction signals: they'd know how happy people actually are with a product, with way more information than a review page.
• Returns on their own: an agent can return stuff for you, with even less human action.
• Safety nets: a cool-down window, insurance for bad purchases, and so on.

Put it all together and you get to present yourself exactly the way you want, with the most optimal consumption possible.