What Agentic Checkout Is, and Why Grocery Got It First

You know what you want for dinner. You have known for an hour. It is 9:20 at night and the cart still has four items in it, because between knowing and owning there are nineteen searches, six quantity judgements and one substitution you will resent.

Agentic AI grocery is the industry's name for deleting that hour. Whether it deserves the name is a fair question.

The word agentic got attached to everything this year, so it is worth being precise. An agent is software that takes an instruction and produces a consequence in the world, rather than producing a suggestion for you to act on. Most of what gets marketed as agentic AI grocery is still firmly in the suggestion business.

The distinction that matters

A recommender ends with a list. An agent ends with a transaction.

That is the whole line. Everything upstream of it, the chat window, the recipe generation, the personalised shelf, is the same technology that has been in these apps for years wearing a new adjective. The genuinely new capability is that software is now allowed to spend your money, with your approval, without you touching a product page.

Suggest is a feature. Buy is a category.

What actually shipped

On 8 December 2025 OpenAI and Instacart launched a grocery shopping app inside ChatGPT, with payment running through the Agentic Commerce Protocol. OpenAI called it the first fully integrated grocery shopping and Instant Checkout app in the assistant, and said Instacart was the first app to offer checkout directly within ChatGPT.

Walmart took the other road. It partnered with OpenAI in October 2025, then in March 2026 dropped Instant Checkout and pushed its own commerce assistant, Sparky, into both ChatGPT and Google Gemini, keeping the transaction on its own rails.

Those are the two strategies the whole of retail is currently choosing between: be a product inside somebody else's assistant, or put your own agent everywhere the customer already is. Read either company's framing as positioning rather than fact, but the plumbing underneath is real and it is new.

Why grocery went first

It looks like an odd category to lead with. Food is personal, perishable and easy to get wrong. Three things made it the obvious testbed anyway.

One. The baskets are enormous and boring.

A grocery order is thirty to sixty line items, most of which you buy every week and have no opinion about. Automating a purchase you have made forty times is low risk and high relief. Nobody is emotionally invested in which brand of foil arrives.

Two. The frequency is unmatched.

You buy a sofa twice a decade. You buy food weekly, which means an agent gets fifty chances a year to be useful, and fifty chances to learn what you meant.

Three. The manual version is genuinely miserable.

Nobody enjoys typing chicken thighs into a search box. The bar for an agent to be an improvement is on the floor, which is a very good place to launch a category that is not yet reliable.

What it still gets wrong

Three failure modes, all of them live, none of them close to solved.

Pantry blindness. The agent does not know what is already in your kitchen. It will buy a second jar of the thing you have most of a jar of, and it will keep doing it, because nothing in the loop tells it otherwise. Every product in the space has this problem and anyone claiming otherwise is reading you a roadmap.

Substitution literalism. Out of stock is where self-serve grocery quietly dies, and the difference between a good agent and a bad one is whether the swap respects what the dish needs. Parsley is not cilantro. Long grain is not arborio. A swap that matches the word instead of the job ruins the meal and you find out at the stove.

Catalogue trust. The agent inherits whatever the retailer's data says. If the store's system claims something is on the shelf and it is not, your agent discovers that at the same moment you do, standing at the pickup bay.

None of this is an argument against the category. It is an argument for keeping the approval step, which is the thing worth checking before you let any of these near a card. We went through it in more detail in the end-to-end breakdown.

The half nobody is agentic about

Here is the strange shape of it. Agentic AI grocery has got very good at the second half of the problem, turning a decision into food in your house. It has done almost nothing about the first half.

You still have to know what you want. The agent is exceptional at executing an instruction and completely silent on the question of what the instruction should be, which is the part that actually stops people at 6PM. Ask any of these tools what is for dinner and you get a recommender again, with all the same problems: too many options, all of them equally plausible, none of them chosen.

So the current state of the art is a chauffeur who will not help you pick a destination. Useful, and only half the trip.

It is worth asking why the industry split it this way, and the answer is not mysterious. Checkout is a solved engineering problem with an obvious owner. Deciding what to eat is a taste problem with no owner, no clean success metric, and a long history of products that answered it by showing you more options and calling that help.

Both halves, one motion

SwipeCart is built on the assumption that deciding and buying are one problem, so treat this as a statement of intent rather than a survey.

You swipe through dinners, crave right and pass left, one at a time with nothing to compare. What you keep goes into a cart, every ingredient matched to a real product at Kroger or Walmart, priced as you swipe. An agent then shops that cart and checks out, after you have approved it and seen the swaps.

The deciding happens when you are hungry, which is when you are honest, and the shopping happens without you, which is when you are useless. If you want the version of this argument aimed at one specific retailer, the Walmart delivery comparison covers it, and the Kroger one covers the other.

One practical note for anyone trying this now rather than reading about it. Start with a reorder, not a dinner party. Give the agent a week you have already bought before, look at what it produces, and pay attention specifically to the substitutions it proposes rather than the items it got right. The right items were never in doubt. The swaps are the whole product, and five minutes of looking at them will tell you more than any review, including this one.

An agent that will not help you decide is a very fast way to buy the wrong dinner.