DoorDash on agentic commerce and DOT (NoPriors)

Notes on the NoPriors interview with DoorDash's two co-founders — a natural-language ordering surface that half the time sends people to a merchant they have never used, a delivery robot built to a use case rather than to a capability, and an internal benchmark built to work out what a 20× rise in model spend actually bought.

Source: a NoPriors episode with Andy Fang and Stanley Tang, DoorDash's co-founders. The transcript arrived as a paste with no URL and no date; its own internal references — a CLI launched "last week", a benchmark announced "a couple weeks ago", June's model spend discussed in the past tense — put the recording in July 2026 at the earliest. Archived under sources/2026-07/doordash-nopriors-agentic-commerce-and-dot/, which also records the transcription damage and which attributions are inferred.

This is the first source in the notebook from a company that buys this tooling at scale and deploys autonomy in the physical world, rather than a lab describing a harness it built or an operator describing his own workflow. That is most of its value here. Two of its threads land on pages that already exist: the argument that the customer is no longer the human gets a vendor actually doing it, and jaggedness gets an instance observed from the buyer's side of a contract.

Ask DoorDash, and the two numbers

The product is a conversational surface over the catalogue. Fang's history of it is worth keeping: they started a couple of years ago and were bullish on voice as the modality, and voice "ended up not being the thing". What worked was the natural conversational experience — people translating what is in their head directly, instead of doing research elsewhere and then keyword-optimising a search box.

Two figures are given for what changed in behaviour:

• Restaurants: 50% of Ask DoorDash trajectories end at a merchant the user has never ordered from before. He flags this as historically one of the hardest metrics at DoorDash to move at all.
• Groceries: about 40% larger baskets. The described use is not search but a task — photograph the fridge and ask it to restock, meal-plan around dietary constraints, or reorder the usuals without tapping through.

The interviewer's read is the interesting part: DoorDash had never seemed hard to use, so the numbers imply latent demand the old interface was not serving — people build habits, but also want variety, and nothing in a search box let them express that.

One deliberate investment: world knowledge. What is trending online about restaurants, outside DoorDash's own data and past the model's cutoff. Fang's stated reason is trust — people want to eat what is current, and an assistant that clearly knows what is being talked about is one they believe.

The agent-first storefront

Asked about five years out, Fang declines and then gives the shape anyway: if college kids started DoorDash today it would look very different, "probably more agentic first". The stat he says he keeps returning to is that there is now more agent traffic on the web than human traffic — no source, no definition — and the question he draws from it is how you have a DoorDash-shaped experience that plays into that.

The concrete move is a DoorDash CLI, launched the week of the recording, described as reducing the friction for an agent to participate. His worked example came from a user who wanted to automate the office-manager job, discovered DoorDash sells more than lunch, pointed a camera at the pantry shelf, and now fires the agent to restock when the shelf looks empty. The same example turns up earlier in Tang's half of the interview, so it is evidently the demo they are both carrying.

That is the vendor side of an argument this notebook already has from the consumer side — see The customer is not the human anymore, where it also answers one of that page's objections and raises a different one.

DOT

Tang's half. DoorDash has been working on autonomy since 2018, starting as "me and half an engineer's time" — a skunkworks whose brief was to partner and learn rather than to build. They worked with the whole field, from sidewalk robots up to robotaxis, and drew three conclusions: autonomy was a question of when rather than if; a great deal of non-autonomy infrastructure has to exist around it (their name for what they built is the autonomous delivery platform — APIs, dispatch, merchant integration, consumer experience); and — the one that made them build hardware — the field builds technology first and looks for a use case afterwards.

The form-factor argument follows from the delivery itself: 3–5 miles, about 15 minutes excluding cook time. A 2 mph sidewalk cooler cannot do it. A 4,000 lb robotaxi is built for carrying people, and a passenger can walk the last half block from where it stops — a burrito cannot. So: roughly 300 lb, 20–25 mph, a bike-lane profile that moves between road and sidewalk. Live in Phoenix, "fully autonomous L4", running deliveries for about two years.

The rest of that thread — the first-and-last-100-feet problem, the edge cases nobody writes down, and the claim about which incumbent data is actually load-bearing — is its own page, because the transferable part is not the vehicle.

Worth recording the fleet position too, since it is the answer to the question everyone asks first. Tang's prediction is more Dashers in ten years, not fewer: the business is growing ~25% a year on 9 million Dashers and 3 billion deliveries, and 5–10× from there cannot come from recruiting half of America. The stance is a multimodal fleet — DOT for suburban strip-mall runs, drones where road infrastructure is poor and the order is light, a human for a multi-step grocery order with stairs — with routing as the thing DoorDash owns and merchants seeing one integration regardless. He also expects cheaper delivery to grow demand, which is the Jevons argument applied to a physical service.

Dashbench, and the 20×

The part of the interview closest to what this notebook is about. Model spend in June was about 20× January's, and when asked directly whether it is still climbing, Fang says it is flatlining — deliberately. Dashbench is their published benchmark over coding tasks, scoring models and harnesses, built to work out the ROI on that spend and to route cheap tasks to open-weight models while paying frontier prices only where they buy something.

And the finding that is worth more than the numbers: ask a team how the models do on their task and the answer is "yeah, it works okay"; scrub the data, hand it to a lab, build an RL environment around it, and the models crush it; run it against real enterprise data with all the real stuff attached and it does not hold. Fang's own framing of the open question — is that a harness gap, or something missing from the models' data distribution — is the honest one, and he does not resolve it.

Both halves are their own page: the ROI question a buyer has to answer, and why the scrubbed-versus-real gap is the same boundary jaggedness draws from the training side.

Since ingesting this, the benchmark itself has been described in public — see dashbench-measuring-a-code-review-agent, published two days before the earliest date this recording can have. It narrows one claim above: DashBench scores a code reviewer on replayed pull requests, not coding tasks in general, and coding agents on real tasks are named there as future work. The "models and harnesses together" reading holds up.

Two smaller things from the same stretch. They acquired a company (transcribed as Metis) last year explicitly to import AI-native working habits, on the diagnosis that a 10,000-person company's problem is not access but that people cannot see what is possible because they are used to how things worked. And seat growth is fastest in the non-technical organisations — analysts, operators, account managers automating merchant QBRs — while the spend is still overwhelmingly engineering. They have no benchmark for that work yet, which is where they say they are going next.

Tasks, and buying the labels

Mentioned in passing and easy to miss: DoorDash launched a product called Tasks a few months before the recording, where people in the Dasher fleet collect data points to help train world models. Fang's own caveat is that this is very early and that the field disagrees about which form factors and model types will work.

It is a real answer to a question left open in the tax-agent notes — what you do when labelled corrections are not a free by-product of somebody's job. You pay a fleet to generate them. That is available to DoorDash and to almost nobody else.

How to read this

Two founders on a podcast, talking about their own products. Nothing here is audited and almost nothing is falsifiable from the outside:

• The behaviour numbers have no denominators. "50% of trajectories" is a ratio over an unstated base of users who self-selected into a new feature, over an unstated window; larger grocery baskets are exactly what you would expect from the subset of customers willing to try a restocking assistant, with or without any causal effect. Both numbers are interesting as directions and worthless as effect sizes.
• "More agent traffic than human traffic" is asserted, not sourced. It is doing real work in the argument, and it is the kind of figure that depends entirely on whether scrapers count.
• "Fully autonomous L4" is a self-report. No disengagement rate, no intervention rate, no incident record, no word on remote assistance, one metro. The comparison to riding a Waymo in San Francisco is theirs.
• The spend story is one-sided. A flatlining cost curve is a cost curve. No output measure is offered against it, so "we need to calculate the ROI" is where the account stops rather than a result.

What survives that discount is the reasoning rather than the evidence: build toward a use case, expect the physical world to be messier than the model, and go looking for the data nobody else has instead of the data everybody has.