Building an Autonomous Delivery Experience with DoorDash Co-Founders Andy Fang and Stanley Tang

2026-07-23 - 50 min - source - Read full transcript

Sarah Guo (host)Andy FangStanley Tang

Key insights

Natural-language ordering surfaces demand that keyword search was suppressing.

50% of Ask DoorDash restaurant trajectories are orders from places the user has never ordered from before, and grocery baskets run about 40% larger than normal - both metrics DoorDash calls historically hard to move. Andy Fang attributes this to conversational search removing the friction of typing keyword queries for things like meal planning or restocking a fridge from a photo.

agentic-commerce

DoorDash speculates that a DoorDash founded today would be built agent-first, not app-first.

Andy Fang argues there is now more agent traffic on the web than human traffic, and that a from-scratch DoorDash would be designed around that reality rather than a human-tapping-through-menus interface - a speculative but deliberate reframing of what the product's core interaction model should be.

agentic-commerce

DoorDash spent years partnering with autonomy startups before concluding it had to build its own robot.

Starting in 2018 as a roughly half-an-engineer skunkworks project, DoorDash worked with both sidewalk-robot and robotaxi companies as a distribution platform. That multi-year partnership phase validated that autonomy was inevitable and taught them the ecosystem autonomy needs (dispatch, merchant integration, operations) - but it also revealed that no partner was building technology use-case-first, which is what ultimately pushed DoorDash to build the technology itself.

autonomous-delivery-robotics

Neither existing robotics category fit DoorDash's actual delivery profile.

Sidewalk robots (2-3 mph) are too slow for the average 3-5 mile DoorDash delivery inside a 15-minute window; robotaxis (4,000 lb vehicles built to carry people) are the wrong form factor for a few burritos. DoorDash concluded the right shape was closer to an autonomous motorcycle or bike - about 300 lbs, up to 20-25 mph, able to travel on both roads and bike lanes - and built Dot to fill that gap since no one else was.

autonomous-delivery-robotics

DoorDash treats autonomy as one modality among several, not an all-or-nothing bet.

Dot is deployed for dense-suburban 3-5 mile trips from strip malls (its current market is Phoenix/Tempe); drones are earmarked for rural areas with poor road infrastructure; and human dashers stay on complex multi-step orders like grocery picking and packing that require climbing stairs. This lets DoorDash phase in autonomy per use case rather than replace one system wholesale.

autonomous-delivery-robotics

The 'first and last hundred feet' of a delivery is a data problem generic mapping can't solve, and DoorDash claims a unique data advantage there.

A GPS pin dropped on an apartment complex doesn't tell a robot which storefront, front door, or gate to use - a human dasher improvises, but a robot can't. DoorDash's advantage is years of historical data on exactly where human dashers actually dropped packages off, which Stanley Tang says doesn't exist anywhere else, including Google Maps.

physical-world-complexity

Deploying autonomy at scale surfaces failure modes invisible in a demo environment.

Examples raised: leaves or dirt covering a sensor changes the torque needed per wheel when only half the wheels are on debris; hard braking can trigger regenerative braking that overloads the battery and causes an electric shock risk; and a robot boot-up script that worked fine for a handful of units took 30-45 minutes and crashed constantly once scaled to hundreds of robots a day.

physical-world-complexity

Real-world edge cases can't be imagined in advance - they have to be discovered by deploying in the physical world at volume.

Andy Fang, referencing his investment in the robotics startup Sunday, gives the example of a cat sitting in someone's dishwasher: no engineer designs for that scenario by imagining it, they only find it (and enough similar cases to matter) by actually operating in enough real homes or streets. This is offered as the core argument for why DoorDash's operational scale and data are a genuine moat, not just a marketing claim.

physical-world-complexity

As autonomy itself becomes less of a bottleneck, operations and hardware manufacturing become the harder scaling problems.

Stanley Tang frames the last five years as autonomy going from an open research question to a solved-enough capability (citing Waymo and Tesla's progress alongside DoorDash's own L4 milestone). The next scaling constraints are operational (adapting to how restaurants differ city to city) and hardware (supply chain, component reliability at fleet scale) - which is why DoorDash partnered with Also, a Rivian spinout, to help scale manufacturing beyond the first hundred hand-built robots.

physical-world-complexity

Internal AI spend at DoorDash grew roughly 20x from January to June before flattening once ROI discipline kicked in.

Andy Fang says the spend flattened after DoorDash started deliberately measuring returns - including via Dashbench, an internal benchmark released to gauge how well different models and harnesses perform on the company's actual coding tasks - and began routing cheaper tasks to open-weight models instead of defaulting to frontier closed-weight models for everything.

ai-productivity-benchmarking

Frontier models look strong on cleaned benchmark data but underperform on raw enterprise data, and it's unclear whether that's a harness gap or a model gap.

DoorDash found that when internal accounting and analytics data was scrubbed and reshaped into an RL-environment-friendly format for a frontier lab, the models performed very well - but performance dropped when tested directly against messy real company data. Andy Fang frames the open question as whether this is fixable with better tooling/harness work, or reflects a genuine gap in what's in the model's training data distribution.

ai-productivity-benchmarking

DoorDash predicts more human dashers in ten years, not fewer, despite investing heavily in robots.

Stanley Tang argues that DoorDash's growth trajectory (roughly 25% year-over-year, with 9+ million dashers already) means that even a 5-10x larger delivery volume in a decade cannot realistically be filled by proportionally more humans - so autonomy has to absorb a growing share of new volume - but he expects total demand to grow even faster than autonomy can substitute for humans, especially as lower delivery costs pull in more orders.

future-of-work-dashers

Companies

DoorDash - Central subject of the interview - framed by its co-founders as a robotics and autonomy company that has been building toward this for eight years, not just a food-delivery app
Waymo - Repeated reference point for what full L4 autonomy looks like, and for the pickup/dropoff constraints (walking half a block) that don't work for package delivery
Tesla - Cited alongside Waymo as another company that has recently made the autonomy 'breakthrough' DoorDash also claims to have hit
Sunday - Robotics (dishwashing) startup Andy Fang has invested in; used as an example of real-world edge cases (a cat in the dishwasher) that can't be imagined in advance
Also - Micro-mobility hardware company spun out of Rivian that DoorDash partnered with to scale manufacturing of the Dot robot beyond the first hundred hand-built units
Rivian - Noted as the company Also spun out of, with Rivian's RJ Scaringe as Also's board founder and chairman
Metis - AI company DoorDash acquired last year specifically to infuse AI-native thinking into the broader organization
McDonald's - Cited as an example of a large-chain merchant integration that behaves very differently operationally from a small independent restaurant
Starbucks - Cited alongside McDonald's as another large-chain merchant integration example with its own operational quirks

Techniques and frameworks

First and last hundred feet problem - Stanley Tang's term for the unsolved precision problem of the exact pickup and dropoff point - which storefront door, which driveway, which gate - that generic mapping data can't answer
Multimodal delivery strategy - Matching each delivery to the modality best suited to it (dashers, Dot, drones) rather than betting on one robot to solve every use case
Dashbench - DoorDash's internal benchmark, released to measure how well AI models and harnesses perform on the company's own coding tasks, used to calculate ROI on internal AI spend
Build toward a use case, not tech-first - DoorDash's stated philosophy, borrowed from their YC 'build something people want' training, of validating a specific customer problem before building general-purpose technology - contrasted with autonomy startups that build the tech first and retrofit a use case

Summary

Sarah Guo sits down with DoorDash co-founders Andy Fang and Stanley Tang to make the case that DoorDash has quietly been a robotics and AI company for most of its life, not a food-delivery app that recently discovered AI. The conversation moves between two of the company's biggest current bets: Ask DoorDash, a natural-language ordering interface, and Dot, an in-house autonomous delivery robot that has been running real deliveries in Phoenix for over two years.

On the agentic commerce side, Andy Fang describes Ask DoorDash's early traction: half of the trajectories where people use it for restaurants are orders from places they've never tried before, and grocery baskets built through natural-language interaction (photographing a fridge, describing a dietary constraint, planning a family dinner) run about 40% larger than normal. He frames this as latent demand that keyword search was suppressing, and speculates that a DoorDash built from scratch today - in a world with more agent traffic than human traffic on the web - would be designed agent-first rather than around a traditional tap-through app.

The robotics story goes back to 2018, when DoorDash started as a roughly half-an-engineer skunkworks project partnering with sidewalk-robot and robotaxi startups rather than building anything itself. That multi-year partnership phase taught them three things: autonomy's arrival was a question of when, not if; a real autonomy ecosystem (dispatch, merchant integration, operations) has to be built around any robot; and, critically, none of the existing autonomy companies were building use-case-first - they built the technology and then went looking for a problem to fit it into, which DoorDash felt consistently produced the wrong product. Neither category on the market fit DoorDash's actual delivery profile: sidewalk robots are too slow for a 3-5 mile delivery inside a 15-minute window, and robotaxis are oversized, overweight vehicles designed to carry people, not a couple of burritos. That gap led DoorDash to design Dot as something closer to an autonomous motorcycle - about 300 lbs, up to 20-25 mph, able to use both roads and bike lanes.

Much of the middle of the conversation is a catalog of physical-world lessons that only appear at scale: leaves or dirt on a sensor changing wheel torque unevenly, hard braking triggering a regenerative-braking overload risk, and a boot-up script that worked fine for a handful of robots taking 30-45 minutes and crashing constantly once scaled to hundreds of units a day. Stanley Tang's broader argument is that DoorDash's ~3 billion annual deliveries and years of data on exactly where human dashers actually drop packages off - not just a GPS pin - constitute a real, hard-to-replicate advantage on what he calls the "first and last hundred feet problem." As autonomy itself becomes less of a bottleneck (both founders point to Waymo and Tesla's recent progress alongside DoorDash's own L4 milestone), the harder constraints going forward shift to operations (restaurants behave differently city to city) and hardware manufacturing at scale - which is why DoorDash partnered with Also, a Rivian spinout, to help industrialize production beyond the first hundred hand-built robots.

The conversation also covers DoorDash's internal AI adoption: engineering AI spend grew roughly 20x from January to June before flattening once the company started measuring ROI more deliberately, including through Dashbench, an internal benchmark for coding-task performance, and by routing cheaper tasks to open-weight models. Andy Fang flags an unresolved tension: frontier models perform very well on internal accounting and analytics tasks once the data has been cleaned and reshaped for evaluation, but underperform on raw, messy enterprise data - and it's unclear whether that gap is fixable with better tooling or reflects something genuinely missing from the models' training distribution.

Closing out, Stanley Tang predicts DoorDash will employ more human dashers in ten years, not fewer, arguing that demand growth (delivery becoming cheaper and more convenient) will outpace how quickly autonomy can substitute for humans, especially given DoorDash's own growth rate. Andy Fang extends the agentic-commerce thread beyond food, citing early users of DoorDash's newly launched CLI who pointed a camera at an office pantry shelf to auto-trigger restocking orders when supplies ran low - offered as a small but concrete signal of what agent-first commerce use cases might look like as the interface gets easier for both humans and agents to use.

Notable Quotes

"If someone were to create DoorDash today, like, I don't know, like college kids in a garage trying to start DoorDash, I think it would look very different, probably more agentic first." - Andy Fang

"If you're only carrying a couple of burritos around, do you really need a 4,000 pound car with chairs and AC?" - Stanley Tang

"If we're trying to do the dishes, why is a cat in the dishwasher?" - Andy Fang

"My prediction actually is in a world where robotics, drones, AI is everywhere, my guess is that in 10 years' time, we're actually going to have more dashers doing deliveries, not less." - Stanley Tang