No Priors

Andy Fang & Stanley Tang on Why DoorDash's Moat Is the Real World

Andy Fang and Stanley Tang· Co-founders of DoorDash at DoorDash
·~49 min·English·No Priors
AgentsRoboticsBusiness Strategy
TL;DR

DoorDash's co-founders explain why they treat the company as an AI and robotics operator, building a custom delivery robot and a natural-language ordering agent from the customer's use case backward and defending it with ten billion deliveries of real-world data.

01The Philosophy

Start From the Use Case, Work Backwards

DoorDash starts every product — ordering agents or delivery robots — from a concrete customer use case and works backward to the technology, the same philosophy it carried from a Stanford dorm-room experiment to autonomous delivery.

It always starts out as what is the customer problem you're solving for? What's the use case you're solving for? Work your way backwards

— Stanley Tang, No Priors
Key Insight
The tell is where each company starts. Autonomy startups that built the technology first, Stanley argues, ended up retrofitting a problem onto it; DoorDash's habit of starting from a single real delivery and working backward is exactly what let it spec a robot no vendor was offering.
02Agentic Commerce

Natural Language Unlocked Latent Demand

Letting people ask DoorDash for food and groceries in plain language surfaced demand the tap-through app never reached: half of restaurant orders through the agent go to places the customer had never tried, and grocery baskets run about 40% larger.

50% of those trajectories are people ordering from places they've never ordered from before, which is huge because that's one of the hardest metrics historically for Door Dash for us to uh move.

— Andy Fang, No Priors
Key Insight
The host's surprise that DoorDash was ever hard to use is the point: the friction was never the catalog, it was translation. A conversational interface let people say what was in their head instead of reverse-engineering it into keywords, and it moved a metric DoorDash had historically found hardest to shift.
03The Agent-First Bet

The Next DoorDash Is Agent-First

Andy's bet is that a DoorDash built today would be agent-first by default, pointing to a web that already carries more agent traffic than human traffic and to early experiments like a DoorDash CLI and a pantry camera that reorders when the shelf runs low.

if someone were to create Door Dash today like I don't know like college kids in a garage trying to start Door Dash I think it would look very different probably more agentic first

— Andy Fang, No Priors
Key Insight
This reframes the interface question. If agents, not people, are the ones browsing and buying, the surface that wins is an API and a command line, not a prettier app screen. The pantry-camera and CLI experiments are small, but they are rehearsals for a world where the customer placing the order is itself software.
04Designing Dot

Not a Sidewalk Robot, Not a Robotaxi

When DoorDash looked for a robot to cover its typical 3-to-5-mile suburban delivery, nothing on the market fit — sidewalk robots were too slow and robotaxis were 4,000-pound machines built to carry people, so it built Dot: a 300-pound, 20-to-25-mph, bike-lane vehicle sized to the job.

the right metaphor is probably a autonomous motorcycle or scooter or bike profile vehicle.

— Stanley Tang, No Priors
Key Insight
The vehicle is a direct readout of the use case. A robotaxi optimizes for carrying a person any distance; a sidewalk robot optimizes for cheapness at walking pace. Sizing instead to a few burritos over three to five miles produces a third thing entirely, which is why an off-the-shelf platform would have meant inheriting someone else's problem definition.
05The Physical World

No Two Deliveries Are the Same

The reason a model-first playbook stalls in delivery is variety: across three billion deliveries a year no two look alike, so a snowy Helsinki drop is nothing like a Dallas drive-thru, and pizza, ice cream, groceries and pharmacy each break differently.

we do what over 3 billion deliveries a year. There are no two deliveries that look the same.

— Stanley Tang, No Priors
Key Insight
This is why the just-make-the-model-and-the-rest-is-secondary mindset fails outside software. The same pattern shows up in DoorDash's own AI work: models that ace a scrubbed, dumbed-down task stumble on the messy enterprise data. The long tail of the real distribution is the problem, not a detail to clean up later.
06The Data Moat

The Data That Lives Only at DoorDash

DoorDash's real advantage is data nobody else holds: ten billion completed deliveries that record where a package actually gets dropped, the first-and-last-100-feet detail a map's approximate address pin can never supply.

that data doesn't exist anywhere else. It doesn't exist in Google Maps.

— Stanley Tang, No Priors
Key Insight
An incumbent-data advantage only matters if it maps to the task. Customer records do not help a delivery robot; knowing where the human Dasher historically left the package does. The deeper lesson echoes a robotics example raised in the conversation: you cannot imagine the real-world distribution in advance, so only the company already operating in the mess can find its odd corners.
07Scaling Autonomy

An Autonomy Business Takes More Than Autonomy

Getting the robot to drive itself turned out to be the easy part; scaling a fleet surfaced a long tail of un-glamorous problems, from a boot-up script that crashed half the time across hundreds of robots to depots, charging and a hardware supply chain that became the real bottleneck.

building autonomy uh business takes more than just autonomy. It's like how do you actually scale something in the real real world?

— Stanley Tang, No Priors
Key Insight
The bottleneck moved. Five years ago the open question was whether delivery autonomy was even possible; now that Waymo-class self-driving is real, the hard part is everything around it — manufacturing durable robots, running depots seven days a week, and handling a wheel that hits leaves and needs different torque. Operations, not the model, is where scale is won or lost.
08The Ten-Year View

More Dashers in Ten Years, Not Fewer

The counterintuitive prediction: even as robots and drones roll out, DoorDash expects more human Dashers a decade from now, not fewer, because a business growing 25% a year and aiming to multiply from nine million Dashers will need every delivery modality it can get.

my guess is that in 10 years time, we're actually going to have more Dashers doing doing deliveries, not less.

— Stanley Tang, No Priors
Key Insight
The framing flips automation's usual story. If autonomy makes each delivery cheaper, demand rises to meet it, and no single modality — Dashers, Dot, drones or sidewalk robots — can absorb a 5-to-10x jump alone. So the endgame Stanley describes is not robots replacing people; it is a multimodal fleet in which every channel, human included, grows.