DoorDash: several hundred agent orders against 993 million in Q3
An analyst on DoorDash's third-quarter call asked Tony Xu how the business changes when people order food through AI agents instead of an app. The CEO answered with a number nobody on Wall Street had: third-party agents have placed "several hundred" orders in total, across the whole time they have been live.
DoorDash booked 993 million orders in the same quarter. That works out to about one agent order for every few million human ones, and it is the first hard measurement of a claim the agent industry has been making all year — that the assistant, not the app, becomes the front door to commerce.
What the number is, and what it isn't
The context around it is a business doing fine. Revenue rose 34% to $4.6 billion, marketplace gross order value climbed 36% to $33.9 billion, net income attributable to common stockholders jumped 91% to $466 million, and adjusted EBITDA rose 41% to $1.06 billion, per DoorDash's results release. Order growth of 28% shrinks to 17% once you strip out Deliveroo, which DoorDash bought last year — worth keeping straight, since the acquired volume flatters the headline.
Xu named the agents doing the ordering: Muse, Instinct, Fo "and others," TechCrunch reported. He framed the strategic question carefully. "You never want to fight gravity," he said, "but at the same time, you don't want to be delusional." His read is that everyday food ordering stays in the app, while agents earn their place on considered, repeatable purchases — the office lunch, the weekly grocery run, the thing you reorder without thinking.
Two things the number does not measure. It counts third-party agents only, so orders placed through DoorDash's own Ask and Text agents, or through the MCP connector it shipped in September, are excluded by definition. And it cannot see an agent that drives a headless browser through the ordinary consumer checkout — to DoorDash's fraud systems, that is just a person with a fast thumb. The real agent-originated volume is somewhat higher than "several hundred." It is still, on any reading, a rounding error.
Why the agents aren't ordering
The honest explanation is that the plumbing is barely built, and the parts that exist were built for someone else.
Ordering food is not a checkout. It is modifiers, sold-out items, substitutions, delivery minimums, fees and a tip, and a merchant that runs out of brisket halfway through the cart. DoorDash's own text-to-order agent shipped with a warning in its own help docs to check the cart before paying, because the agent can be wrong. A third-party assistant inheriting that job has to get all of it right, in a conversation, without the interface affordances that let a human fix a mistake in two taps.
Then there is the connector problem. DoorDash published a Model Context Protocol connector that lets an assistant search items, build a cart, place and track an order — but as a limited beta aimed at companies ordering for their own staff, with SpaceXAI, Vercel, Cognition, Tempo and Mercor in the early pool, as we detailed when DoorDash opened its text-ordering waitlist. That is a B2B on-ramp. A consumer agent like Muse has to reach the same endpoints through terms nobody has published, or fake being a human.
Latency does the rest. A reasoning loop that takes 40 seconds to decide between two Thai places is competing with an app that opens on the last order and checks out in fifteen. Hunger has a clock, and agents are still slow.
Who wins if it stays small — and if it doesn't
DoorDash wins either way, which is why Xu can afford to be candid. If agents stay marginal, nothing changes and the company keeps its ad inventory, its DashPass attach rate and its customer relationship. If agents take off, DoorDash wants to be the logistics layer every assistant has to call — the only party that knows what is in stock, manages the merchant handoff and eats the refund when the wrong noodles arrive. That is the same argument it made in August, when Bank of America's Justin Post asked about agentic traffic and Xu said the volume was "quite low" and that consumers care about getting the burrito, not the label on the software that ordered it.
The contrast with Walmart is the interesting part. Walmart spent October rolling out the welcome mat — an agent gateway, per-agent profiles with spend caps, published terms — while Amazon has spent a year blocking agents outright. We went through the mechanics of Walmart's Agent Gateway when it was announced. DoorDash is closer to Walmart's position than Amazon's, and the reason is commercial: it has no first-party storefront to defend, only a fee to collect.
The losers, if the number stays this small for another two quarters, are the consumer agent startups selling the story that commerce is about to route around the apps. Muse, Instinct and Fo have no food distribution of their own. Meta can afford to wait for Muse to matter; a seed-stage shopping agent cannot.
The skeptic's read
"Several hundred" is not zero, and the agents being measured are weeks old. Muse launched on September 8. Judging agentic commerce by orders placed in its first seven weeks is like judging mobile commerce in 2009 by WAP sites.
There is also a motive to talk the number down. DoorDash is negotiating with agent makers over who pays for the order, and a public "it's tiny" is a useful opening bid. Xu even volunteered where the money could come from: advertising inside the agents themselves. Read that as a tell. The moment an assistant picks the restaurant, DoorDash's sponsored placement loses its audience — so the company's answer is to sell the slot inside the assistant instead. That is not a defense of the app; it is a hedge against it.
Meanwhile the quarter itself shows what DoorDash is actually spending its attention on: an agreement to buy Instacart for $65 a share, at least three-quarters in cash. Grocery is where agent ordering has a real use case — repeatable, list-driven, low-impulse — which makes the deal a bet on the version of agentic commerce that arrives after the novelty wears off.
Would you hand a shopping agent your card and let it pick dinner? Write to us via our contact page.
What to watch
The fourth-quarter call. If Xu gives a number again and it has moved by an order of magnitude, this was a launch artifact; if it is still "several hundred," the agents have a product problem, not a distribution one. Then Walmart's consumer agent profiles, due in November, which will produce the first large-scale test of whether shoppers trust an assistant with a spend limit. And whether any agent maker signs a revenue-share with a delivery platform — the first such deal will set the take rate for the whole channel.
