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

Summary

Andy Fang and Stanley Tang argue that DoorDash is being reshaped by natural-language commerce and autonomy rather than just faster delivery. Ask DoorDash is already nudging users toward new restaurants and larger grocery baskets, and the company has started incorporating broader world knowledge so recommendations feel timely and trustworthy. On the robotics side, they describe a multi-year path from skunkworks partnerships to Dot, DoorDash’s in-house autonomous delivery robot, built specifically for 3- to 5-mile deliveries in dense suburbs. The conversation emphasizes that scaling autonomy is less about a model demo and more about operational reality: boot times, sensor contamination, depot workflows, manufacturing, merchant integration, and the unique first/last-100-feet problem. Their broader thesis is multimodal rather than replacement-focused: more autonomy, more drones, and still more humans as DoorDash expands.

Notes

• DoorDash is positioning itself around agentic commerce, not just traditional search and checkout.
• Ask DoorDash is described as a natural-language interface for food and groceries.
• DoorDash says Ask DoorDash is changing restaurant discovery by surfacing novelty.
• The company says Ask DoorDash is driving about 40% larger grocery baskets.
• Users are using the interface for fridge stock-up, meal planning, dietary constraints, and reorder flows.
• DoorDash explicitly added world knowledge beyond model cutoffs and marketplace data.
• The goal of that added context is to make recommendations feel more trustworthy and relevant.
• Andy Fang says a future DoorDash would look more agentic-first if created today.
• Andy Fang says agent traffic on the web now exceeds human traffic.
• A pantry camera can trigger a replenishment query when shelves run low.
• DoorDash says it has treated robotics and autonomy as strategic since 2018.
• Stanley Tang says the next DoorDash competitor will be more than a prettier UI.
• DoorDash’s investment thesis spans AI, agent commerce, autonomy, robotics, and drone delivery.
• The robotics effort started as a skunkworks project with minimal staffing.
• DoorDash initially expected to partner with outside robotics vendors rather than build in-house.
• The team worked with multiple autonomy vendors, from sidewalk robots to robo-taxis.
• DoorDash concluded autonomy needs surrounding infrastructure, ecosystem, and product integration.
• The company built an autonomous delivery platform for APIs, dispatch, merchant integration, and consumer experience.
• Stanley Tang says existing robot platforms were often technology-first rather than use-case-first.
• Sidewalk robots were too slow for DoorDash’s 3- to 5-mile delivery pattern.
• Robotaxis were seen as overbuilt for carrying a few burritos.
• Delivery robotics has a distinct first- and last-hundred-feet problem.
• DoorDash thinks the right vehicle profile is closer to an autonomous motorcycle or scooter.
• The robot was designed using insights from DoorDash’s delivery history.
• DoorDash says it has 10 billion deliveries of internal data and 40 million monthly consumers.
• No two of DoorDash’s 3 billion annual deliveries are the same, according to Stanley Tang.
• DoorDash says the product can also collect training data for world models through Tasks.
• DoorDash dot is an in-house autonomous delivery robot built for sidewalks, bike lanes, and roads.
• The team says it had to collect realistic real-world data rather than rely on demos.
• DoorDash’s modality strategy is to avoid solving every use case with one system.
• The company’s examples include robots for dense suburbs, drones for light rural orders, and Dashers for complex grocery jobs.
• Merchant and consumer experiences are meant to stay unified through one integration.
• DoorDash describes itself as an ecosystem for local commerce.
• Hiring is framed as a chance to ship real products, not just demos or lab work.
• Stanley Tang says scaling autonomy requires more than models.
• Real-world edge cases include dirty sensors, leaf-induced torque changes, braking issues, and boot failures.
• The boot-up workflow originally relied on a hacked Jenkins script that did not scale.
• Scaling also forces work on reliability, manufacturing supply chain, depot operations, recharge, and localization.
• DoorDash says its drop-off data comes from human Dasher behavior and is not available in Google Maps.
• Andy Fang says operations expertise is a critical complement to technology.
• DoorDash says Phoenix autonomy reached L4 last year after scaling from a few robots to about 100.
• The autonomy stack is purpose-built in house, not copied from Waymo.
• DoorDash says the bottleneck is increasingly operations, not autonomy.
• Hardware is becoming a scaling bottleneck because the company must move from hand-built robots to supply-chain-reliable manufacturing.
• DoorDash partnered with Also to learn how to scale vehicles and manufacturing.
• Stanley Tang says the robotics problem has shifted from autonomy to commercialization and hardware manufacturing.
• Andy Fang says DoorDash acquired Metis to bring in AI-native ways of working.
• DoorDash is trying to expose the organization to frontier practices.
• Coding is the clearest AI transformation area, but DoorDash wants AI across the company.
• Dash Bench was launched to benchmark models and harnesses on coding tasks and estimate ROI.
• Andy Fang says AI spend rose about 20x from January to June.
• DoorDash says AI spend is now flattening after an experimentation spike.
• Nontechnical teams are seeing the fastest AI seat growth.
• Model performance on enterprise tasks improves after data scrubbing and RL-environment setup.
• Stanley Tang predicts more Dashers in ten years, not fewer.
• DoorDash expects a multimodal fleet of humans, drones, robots, and other autonomous vehicles.
• Andy Fang says people are increasingly comfortable talking to agents like humans.
• DoorDash is testing agent-first workflows with DoorDash CLI and pantry restocking.