No PriorsBuilding an Autonomous Delivery Experience with DoorDash Co-Founders Andy Fang and Stanley Tang
CHAPTERS
- 0:05 – 1:51
Ask DoorDash: natural-language ordering and why it beat voice (for now)
Sarah opens by probing DoorDash’s push into agentic commerce and the origins of Ask DoorDash. Andy explains they initially bet on voice, but the real adoption came from conversational text that lets people translate intent directly into actions without “keywording” their way through the app.
- •Ask DoorDash began as a multi-year effort to change how users search and order
- •Voice was an early hypothesis, but conversational natural language landed first
- •Users prefer expressing nuanced intent over browsing and keyword optimization
- •Early traction held up as rollout expanded
- 1:51 – 3:25
Behavior change: new restaurants and bigger grocery baskets
Andy shares concrete metrics showing Ask DoorDash changes what people buy. The feature drives meaningful restaurant discovery and significantly increases grocery basket sizes by making planning and restocking easier than tapping through menus.
- •~50% of Ask DoorDash restaurant journeys lead to ordering from a new-to-them restaurant
- •Grocery baskets are ~40% larger when using Ask DoorDash
- •Use cases: photo-of-fridge restocking, meal planning with constraints, reordering “usuals”
- •Conversational UX surfaces latent demand for variety and discovery
- 3:25 – 4:26
Trust and discovery: adding “world knowledge” and making the feature less intimidating
They discuss how DoorDash improves recommendations by incorporating signals beyond its own marketplace. Andy also notes the near-term product challenge is helping users discover what to ask and how to start, since an open-ended prompt can feel intimidating.
- •Incorporating external trends to reduce knowledge-cutoff issues and boost relevance
- •“What’s cool/trending” becomes part of the discovery experience
- •Near-term roadmap: UX experiments to guide first-time queries and use-case discovery
- •Longer-term: DoorDash built today would likely be more agentic-first
- 4:26 – 5:56
Agentic futures: ambient context, office pantry cameras, and agent traffic on the web
Sarah pushes on what richer context could unlock. Andy describes an “ambient” agentic workflow—like a camera watching an office pantry shelf—and notes the broader shift toward agents being major web actors, implying DoorDash must support agent-first interfaces.
- •Example: camera detects low inventory and triggers DoorDash restock automatically
- •Early experimentation includes making DoorDash easier for agents to interface with (CLI)
- •Observation: agent traffic on the web is surpassing human traffic
- •Agentic DoorDash implies new surfaces beyond the traditional consumer UI
- 5:56 – 7:01
From family dinner to office lunch: autopilot ordering as an agentic workflow
Sarah shares a real recurring group-ordering scenario (family dinner with allergies and preferences). Stanley maps it to the similar “office manager lunch” workflow, reinforcing the practical value of agents in coordinating people, constraints, and timing.
- •Recurring group orders involve attendance, dietary constraints, and preferences
- •Autopilot ordering is positioned as achievable with agentic commerce
- •Parallel use case: office lunch ordering with timing and preference management
- •Agentic workflows reduce coordination overhead, not just search friction
- 7:01 – 8:53
Why DoorDash started investing in autonomy early (2018) and how bets get sequenced
Stanley explains DoorDash began exploring robotics and autonomy in 2018 to avoid future disruption. He frames long-horizon bets through DoorDash’s broader experimentation culture: start small, validate, then scale investment as evidence accumulates.
- •Robotics/autonomy exploration started in 2018, before it was “obvious”
- •Founder-led view: next-generation competitors won’t win with just a better UI
- •Capital allocation approach: skunkworks experiments that expand with validation
- •Early years focused on partnerships and learning rather than building hardware
- 8:53 – 17:46
Partnership learnings and the shift to building Dot in-house (use-case-first)
DoorDash worked with many autonomy providers and learned what it takes to operationalize autonomy as a service. The decisive insight: most robotics companies built technology in a vacuum, while DoorDash needed a vehicle designed around its specific delivery use case.
- •Autonomy is “when, not if,” but requires a surrounding ecosystem to work
- •DoorDash built an “autonomous delivery platform” (APIs, dispatch, integrations)
- •Key lesson: technology-first robotics often misses the actual customer problem
- •Sidewalk bots are too slow; robotaxis solve a different pickup/drop-off problem
- •DoorDash chose an in-between form factor (scooter/motorcycle-like) and built it
- 17:46 – 21:20
Physical-world complexity: why demos fail and DoorDash’s data advantage matters
The conversation turns to why robotics is harder than software: the environment distribution is vast and messy. DoorDash’s scale—billions of deliveries—provides real operational and location data (like true drop-off points) that others can’t easily replicate.
- •Every delivery is different across geographies, merchants, and item types
- •Realistic data comes from real-world contact, not cherry-picked demos
- •Unique advantage: historical pickup/drop-off behavior (first/last 100 feet)
- •Multimodal vision: not one robot for everything, but matching modalities to jobs
- •Single consumer/merchant interface while routing to humans, robots, drones, etc.
- 21:20 – 26:18
What Dot is: specs, deployment in Phoenix, and multimodal delivery strategy
Stanley describes DoorDash Dot: an in-house L4 autonomous delivery robot operating in Phoenix/Tempe. They position it as a road-and-bike-lane-capable vehicle designed for 3–5 mile suburban deliveries, fitting alongside drones and human Dashers.
- •Dot: ~300 lbs, up to ~20 mph, ~1/10th the size of a car
- •Designed to operate on bike lanes and roads (not just sidewalks)
- •Live in Phoenix for ~2 years; fully autonomous L4 milestone achieved
- •Strategy: assign deliveries by modality (Dot vs drone vs Dasher) based on task fit
- 26:18 – 35:21
Scaling autonomy: the unglamorous edge cases, operations, and fleet realities
Stanley details why scaling is much harder than proving autonomy works once. Operating daily at scale exposes edge cases (sensor dirt, traction differences from leaves), as well as operational necessities like depots, maintenance, charging, and reliable boot-up workflows.
- •Edge cases emerge only under continuous real-world operation
- •Examples: camera sensor dirt, mixed-surface traction affecting wheel torque
- •Safety-driven extremes: hard braking interacts with regen braking and battery systems
- •Non-autonomy blockers: depots, maintenance, charging, fleet boot-up and reliability
- •Mapping/pinning issues: robots need precise storefront/door/gate-level destinations
- 35:21 – 39:30
From Phoenix to many cities: autonomy, ops interfaces, and manufacturing at scale
Sarah asks what’s next from Phoenix scaling. Stanley breaks constraints into three parts—autonomy, operations/integrations, and hardware/manufacturing—arguing autonomy is becoming less of the bottleneck while commercialization and vehicle scaling become dominant.
- •Three scale challenges: autonomy, ops/interface + fleet management, and hardware
- •Autonomy is increasingly less limiting; new cities introduce new edge cases
- •Operations differ across markets (Phoenix vs SF vs London vs Helsinki)
- •Hardware is not a commodity at 1,000–10,000 units: supply chain and durability matter
- •Partnership with Also (micro-mobility spinoff from Rivian) to help scale vehicles
- 39:30 – 44:55
AI productivity inside DoorDash: spend, benchmarks (DashBench), and ROI questions
The discussion shifts to internal productivity in a 10k+ person company. Stanley and Andy describe infusing AI-native practices (including the Metis acquisition), monitoring rapidly rising model spend, and creating DashBench to quantify coding performance and ROI—while noting harder unsolved benchmarking for non-coding work.
- •Goal: bring frontier AI-native operating practices into a large org (Metis acquisition)
- •LLM spend jumped ~20x from January to June; later stabilized via controls
- •DashBench: benchmark to measure coding-task performance and evaluate ROI
- •Exploring cheaper/open-weight models for some tasks while preserving quality
- •Non-technical adoption is growing (analysts, ops, account managers) but harder to benchmark
- •Real enterprise data/harnessing remains a bottleneck vs “scrubbed” lab settings
- 44:55 – 49:19
The future: more Dashers + multimodal fleets, and agent-first commerce beyond food
Sarah asks whether robots replace humans; Stanley predicts the opposite: more Dashers alongside more autonomy as demand grows and new modalities expand supply. Andy closes by connecting agentic commerce to longer, more natural queries and agent-first experiences (including CLI and ambient restocking) that extend beyond food into broader local commerce.
- •Prediction: in 10 years, more Dashers—not fewer—due to growth and supply needs
- •DoorDash scale: ~9M Dashers, ~25% YoY growth; autonomy helps meet demand
- •Delivery becomes cheaper → demand increases → need every modality available
- •Trend: search queries getting longer; users interact with apps more like humans
- •Agent-first design and CLI enable new workflows (e.g., pantry camera restocking)
- •Closing reflections on ambition across UX (agents) and physical fulfillment (autonomy)
