DoorDash's AI agent orders stay marginal at several hundred
- DoorDash has received only several hundred orders from third-party AI agents, according to chief executive Tony Xu.
- The company has cited an average of about 3 million deliveries a day in 2025.
- Xu named Muse, Instinct and Fo as agents placing orders and said most customers still use DoorDash's own app.
- Amazon has blocked personal agents while Walmart has welcomed them, with no common standard for agent access.
DoorDash has taken only “several hundred” orders placed by third-party AI agents, its chief executive Tony Xu said on the company’s third-quarter earnings call. For scale, the company has cited an average of roughly 3 million deliveries a day in 2025. Set against that, agent-initiated ordering is not a market. It is a test case.
Xu named Muse, Instinct and Fo as the agents involved. He called agentic ordering an area worth following rather than resisting, and warned against overstating its present role. Most customers, he said, still order through DoorDash’s own application. Even so, the company plans to study agentic use cases, build relationships with third-party agent developers, and consider selling advertising inside those agents. Its own work points the same way: a text-to-order agent in September, a command-line agent in July, a chatbot called Ask DoorDash, released in June, which accepts prompts and photographs, and a Model Context Protocol interface that lets other chatbots place corporate orders.
Access on whose terms
The wider argument is about who controls access to a shopfront. Amazon has blocked personal agents. Walmart has welcomed them as a way of drawing customers from rivals. Both positions are commercial rather than technical, and neither rests on a published standard for how an agent should identify itself, what it may do on a customer’s behalf, or how errors are put right. That gap matters. An agent ordering food handles an address, a payment method and a set of substitutions, and the limits of its authority are rarely written anywhere a customer would look.
Advertising inside agents sharpens the point. If an agent ranks restaurants on a customer’s behalf, the ranking becomes a paid placement. The person behind the order may have no way to see why one option came first.
The provenance question
A quieter parallel sits in education, where the same difficulty appears in another setting: a record that shows who acted, on whose instruction, and what was verified afterwards. Coursework and assessment policies are being rewritten around that question, largely without settled answers. Agent commerce will need something similar, since a delivery that goes wrong raises a question about responsibility that a receipt does not answer.
Order volume will be reported every quarter. The rules governing what an agent may decide, and what the person behind it is entitled to see, are being settled far more quietly.
TechCrunch’s report on DoorDash said it’s only received ‘several hundred’ orders from AI agents provides the source reporting for this article.