The next customer is an agent

DoorDash shipped a command-line tool for ordering food this week. It is called dd-cli, it is in limited beta, and it lets an AI agent search restaurants, apply deals, and check out on your behalf. Andy Fang, the company’s co-founder, announced it in a post that passed two million views in its first hours.

A CLI for burritos sounds like a developer toy. I read it as the clearest signal yet of a bigger shift: the customer that consumer companies build for is about to stop being a person with a phone, and start being that person’s agent.

Strategy, not a stunt

It would be easy to file dd-cli under marketing. The paper trail says otherwise. DoorDash acquired Metis, a YC-backed applied AI lab, in March 2026 to accelerate what it calls agentic commerce. Months before that, Fang wrote that DoorDash would need to “earn the right to service customers (agents?) end-to-end.” That parenthetical question mark is the whole story: the company is re-deciding who its customer is.

The beta’s shape backs this up. US and Canadian macOS developers, waitlist only. That is not a consumer launch. It is an invitation to the people who build agents to start treating DoorDash as a callable service.

The layer that disappears

Everything you order through today was designed for human eyes and thumbs: screens to browse, photos to scroll, buttons to tap. An agent needs none of it. It needs a surface it can query, compare, and transact against.

Play the scenario forward. You tell your agent you’re hungry. It knows you had pasta yesterday, that you’re allergic to peanuts, that it’s the end of the month and the budget is tight, and that your usual place quotes a 40-minute wait tonight. Fifteen minutes later food is at your door, and you never opened an app.

And you did not type any of that. You said it. The typed CLI is the developer preview; the version your parents use is a conversation. Part of it already ships: “Alexa, reorder the dog food” has completed real purchases by voice for years, and in 2026 Amazon’s shopping agent gained the memory to resolve “my usual,” weigh a budget you set out loud, and check out on stores Amazon does not even own. The full dinner version, allergies and wait times folded in, is the near future rather than today, and the honest reason is telling: OpenAI quietly walked back in-chat checkout in March, and DoorDash retired its own restaurant voice bot around April. The pieces work. Stitching them into something you would trust with a peanut allergy is the hard part, which is the whole point of this piece.

Notice what vanished from that story: browsing, search results, the featured carousel, the promoted listing. The entire discovery layer, the one consumer companies spend billions fighting over, collapsed into an API call. When the customer is an agent, discovery, brand loyalty, and pricing all get renegotiated in machine-readable terms.

The race is already on

DoorDash is not early to an empty field. Stripe and OpenAI co-developed the Agentic Commerce Protocol and shipped Instant Checkout inside ChatGPT. Visa is building Intelligent Commerce, Mastercard has Agent Pay, and Google leads the AP2 coalition with PayPal and Amex. The payment networks are not debating whether agents will transact. They are fighting over how.

The sharper signal sits on GitHub. Before DoorDash shipped anything official, developers had already published at least four unofficial DoorDash MCP servers, some reverse-engineering the app’s internal APIs so their agents could order food. Demand preceded supply. When developers start picking your locks to integrate with you, shipping them a door is overdue.

Why you should ship one

If an agent can consume your product, an agent-native surface is distribution. Agents will settle into defaults the way people did: the service that answers reliably, returns clean structured data, and doesn’t throw up a CAPTCHA becomes the one that gets called first. Holding that position early looks like ranking first in search in 2005, or being in the App Store in 2010. Those land-grabs weren’t won by the best products. They were won by the products that took the new channel seriously first.

The same logic lands harder in markets where super-apps already own the wallet. If I ran platform engineering at a Careem or a Talabat, an agent surface would be on this quarter’s roadmap, not next year’s.

To be precise about the claim: apps are not dying next year. Retail giants are deliberately keeping the transaction inside their own properties, and browsing will survive wherever browsing is the pleasure; nobody wants an agent choosing their art. The shift is narrower, and still enormous. For every purchase that is a chore rather than a joy, the interface of record moves to the agent, and the strategic surface of the business moves with it.

The boring part decides whether it sticks

Agent traffic does not behave like human traffic. It retries at machine speed. It never hesitates before confirm. And it fails in a way no payment client before it ever did: an agent that times out does not just resend the same request. It can re-plan, rebuild the cart from scratch, and submit what looks like a brand-new order. I have spent most of my career on this class of problem, from payment systems at MNT-Halan (I joined as one of the first ten engineers; it became Egypt’s first fintech unicorn) to an open-source double-entry ledger I built this year. The classic defense is the idempotency key: a retried request carries the same key, and the second attempt returns the first result instead of charging twice. That defense still works, and it is structurally blind to the re-planned order, which arrives with a fresh key and sails through deduplication as a legitimate new sale. I take that mechanism apart, code included, in its own deep-dive.

The payment protocols underneath this wave have done serious work here, more than the launch posts advertise. The Stripe and OpenAI checkout protocol now makes idempotency keys mandatory on every call, and caps each purchase with a single-use token bound to one merchant, one amount, one expiry. Google’s AP2 skips request keys entirely and builds on signed mandates: the user’s actual instructions, budget, and recurrence limits, carried as cryptographic proof. Visa and Mastercard register purchase intent with the network and check each authorization against it. One pattern runs through all of it: the old layer dedupes identical retries, and a new layer caps the damage from everything the old layer cannot see. Google’s own spec still names duplicate orders across checkouts as an open problem.

The checklist for anyone shipping an agent surface follows the same pattern, and it is unglamorous. Idempotency keys on every call that moves money. Authorization holds and scoped single-use grants instead of an open credential. Per-agent spend caps. An audit trail a human can read the next morning, because the failure will not happen on the happy path; it will be the duplicate charge at 3 a.m., the allergy field that didn’t sync, the refund nobody can trace. Online card payments became universal only after the trust rails (chargebacks, 3-D Secure, PCI) made failure survivable. Agent commerce will follow the same curve, and the companies that ship trust rails with their API will be the ones agents are still allowed to call next year.

Who will your agent order through?

Two million people watched a founder announce a command-line interface for food. Whether agents will buy on our behalf is no longer the question. The open question is which services they will buy through, and whether yours will be one they can trust with a wallet.

I’m hungry, and soon my agent will handle that end to end. The only question is whose API it calls.