DoorDash has put a number on agentic commerce, and it is small. On the company's third-quarter 2026 earnings call, CEO Tony Xu said DoorDash is seeing "maybe several hundred orders in total" from third-party AI agents, according to TechCrunch's report from October 30, 2026. Set that beside the roughly 3 million deliveries a day in DoorDash's 2025 report and the gap is plain: agents are an experiment at DoorDash, not yet a channel.
This post explains what was said, what is and is not confirmed, why the figure matters for anyone building agents that buy things, and how it fits the rest of the agent-commerce story we have been tracking on explainx.ai.
TL;DR: what was said and what it means
| Question | Short answer |
|---|---|
| What did DoorDash say? | Third-party agents have sent "maybe several hundred orders in total" (CEO Tony Xu, Q3 2026 call) |
| Where was it said? | On the Q3 earnings call, reported by TechCrunch |
| Which agents? | TechCrunch names Muse, Instinct and Fo as examples, plus unnamed others |
| Exact count? | Not given. "Several hundred" is Xu's own approximation |
| Is DoorDash against agents? | No. It is building its own agent surfaces and wants relationships with agent makers |
| Biggest takeaway | Demand from third-party agents is real but tiny; most people still open the app |
What Tony Xu actually said
According to TechCrunch, Xu said you "never want to fight gravity," but that you also should not be "delusional," and then gave the figure: "we're seeing maybe several hundred orders in total from some of these third-party agents."
Two things are worth noticing. First, the number is cumulative ("in total"), not per day. Second, it is a rounded verbal estimate from an earnings call, not a filed metric. Treat it as an order-of-magnitude statement: hundreds, not thousands, and certainly not millions. The earnings materials are on DoorDash's investor relations site if you want to check what the company files versus what executives say aloud.
TechCrunch also reports that most users still order through DoorDash's own app, that DoorDash wants to learn about agentic use cases and build good relationships with third-party agent makers, and that it may eventually be able to advertise inside these agents. That last point is the business logic: if agents become the front door, DoorDash wants to be present and monetizable there.
How big is "several hundred"?
Using the figures TechCrunch cites, 3 million deliveries a day is about 90 million a month. Several hundred orders in total is on the order of a millionth of one month's volume, and that is before counting that the agent orders accumulated over an unspecified, longer window. Exact arithmetic is impossible because DoorDash did not give an exact count or a time window, so do not quote a percentage as a company figure. What can be said safely is that the share is negligible today.
That is not a verdict on the future. Early channels are always small. The useful question is what has to change for the number to grow, and the answer is mostly not about DoorDash.
DoorDash is building its own agents anyway
The "agents are small" story sits beside a lot of DoorDash agent activity. Per TechCrunch, the company launched:
- Ask DoorDash in June, a chatbot that takes orders from prompts and photos.
- A command-line agent in July, which we covered in DoorDash dd-cli: order food from your AI agent in the terminal.
- A text-to-order agent in September.
- An MCP connector that lets other chatbots place orders, aimed at corporate ordering.
The pattern is deliberate. DoorDash controls the first-party surfaces, where it keeps the customer relationship, the ads and the data, and offers a connector so outside agents can plug in. Our write-up of DoorDash Flux and its cloud agents platform shows the same instinct on the engineering side: use agents heavily internally, expose them carefully outside.
Why third-party agents are not ordering much yet
The earnings call did not explain the low number, so what follows is our reading, not DoorDash's. Several plausible frictions stack up:
- Trust at checkout. Letting an agent spend money needs spend caps, confirmation steps and a clear way to dispute a wrong order. Most people are not ready to hand that over for dinner.
- Convenience. For a repeat order, the DoorDash app is already two taps. An agent has to beat that, and a chat round trip often does not.
- Access rules are unsettled. Across the industry, websites are still deciding whether to let agents in at all. TechCrunch notes that Amazon blocks personal agents while Walmart has welcomed them. We covered both sides in Walmart's agent gateway with spending limits and Amazon's blocks and in Paul Graham on Amazon banning agents.
- Few agents with real distribution. The named agents are small. The big assistants have not made food ordering a headline feature, so there are few users in the funnel.
- No common standard yet. Payments and identity for agents are being worked out in several places at once; see the personal agent protocol from Sierra, Meta, Walmart and Stripe and Stripe Link and agentic commerce.
What this means if you build agents
A few practical takeaways follow from a number this small.
Do not size a business on agent-originated orders. If your plan assumes consumers will route most purchases through a third-party assistant within a year, DoorDash's data point says that is not happening yet at one of the largest delivery platforms.
Pick narrow cases where an agent beats an app. The MCP connector is aimed at corporate ordering for a reason: a group order for a team lunch, with a budget and an approver, is more tedious in an app than in a chat. Repeated, rule-bound, multi-person purchases are where an agent saves real effort.
Design the permission layer first. The same trust gap that limits adoption is the thing you can fix. Cap spend, require confirmation above a threshold, log every action and make orders reversible. If you are wiring agents to real money, the AgentBeam platform from the explainx.ai team is built to stop agents before they take dangerous actions, which includes risky purchases.
Expect platforms to want the relationship. DoorDash said it may advertise inside agents. Whoever owns the agent will be asked to share ordering data and ad slots. Negotiate that early.
A checklist for testing agent ordering yourself
If you want to find out whether agent ordering works for your own use case, run a small, controlled test rather than a launch.
- Start with a fixed budget. Give the agent a hard cap per order and per day, and use a payment method that cannot overspend.
- Use a repeatable task. A recurring team lunch or a standing grocery list lets you compare the agent's result with what a person would have done.
- Measure the right things. Track time saved, error rate (wrong item, wrong address, duplicate order), and how often a human had to step in.
- Keep a human confirmation step until the error rate is boringly low. Autonomy is a dial, not a switch.
- Read the terms. Platforms differ on whether automated ordering is allowed, so check the access rules before you build on a connector.
The outcome of a test like this is more useful than any headline number, including the one from the earnings call. It tells you whether the friction is the agent, the platform, or the users.
What is verified and what is not
| Claim | Status |
|---|---|
| Xu said "maybe several hundred orders in total" | Reported by TechCrunch from the earnings call; we have not seen a filed metric |
| About 3 million deliveries a day | From DoorDash's 2025 report, per TechCrunch |
| Muse, Instinct, Fo are ordering | Named by TechCrunch as examples; volumes not disclosed |
| Ask DoorDash, CLI, text-to-order, MCP connector exist | Listed by TechCrunch with launch months |
| Reasons agents are low | Our analysis, not stated by DoorDash |
If you want the exact wording, find the Q3 2026 call transcript or recording on DoorDash's investor site and check the answer yourself. We have quoted only what TechCrunch attributes to Xu.
The bigger picture
Agentic commerce has been one of the most hyped ideas of 2026, with protocols, payment rails and launches from several large companies. A candid CEO saying the real volume is "several hundred orders" is a useful dose of calibration. It does not mean the idea fails. It means the infrastructure got built faster than the habit did, and that habits, not protocols, set the pace.
Watch three signals over the next few quarters: whether DoorDash gives a firmer number, whether a large assistant makes ordering a default feature, and whether platforms like Amazon change their stance on agents. If any of those move, the several hundred could become several hundred thousand quickly. Until then, the honest description is early.
Related reading
- DoorDash dd-cli: order food from your AI agent
- DoorDash Flux: cloud agents platform
- Walmart agent gateway, MCP and Amazon blocking agents
- Paul Graham on Amazon banning agents
- Personal agent protocol: Sierra, Meta, Walmart, Stripe
- Agentic commerce with Stripe Link
Figures and quotes are as reported on October 30, 2026 and may change if DoorDash publishes more detail.