Ecommerce ops · Production

DoorDash Assistant: engineering overview of a conversational agentic shopping assistant

The problem

Local-commerce grounding data — menus, prices, hours, ETAs, delivery radii, and inventory — changes minute to minute and does not live in a model's weights, making reliable grounding of a conversational shopping agent a core engineering challenge.

First attempt

Early production failures included the agent recommending stores as open when they were closed, citing prices that did not match the catalog, and claiming to have added items that were not actually in the cart.

Workflow diagram · grounded in source

Consumer submits request

trigger

“A consumer describes what they want, for example "a quick dinner under $30 near me" or "this week's groceries for two people, vegetarian, $60 budget," and the Assistant produces a response the consumer can refine through the conversation”

Consumer memory retrieval

ai_action

“The agent retrieves the consumer's memory blocks (dietary preferences, brand affinities, past order history)”

Live catalog and store search

integration

“runs a delivery-radius search for currently open grocery stores with reasonable ETAs, picks one, and assembles a shopping list”

LLM orchestration assembles response

ai_action

“6-8 LLM calls and a handful of tool calls against the live catalog (consumer memory lookup, store search, per-merchant inspection, item search, order history lookup, optional pricing or deals check, display planning, reply text + suggest…”

Grounding validation via tool call

validation

“The fix in each case has been to route the agent's claim through a tool call against the system of record”

Widget rendered to consumer

output

“The Assistant renders the list as an interactive widget with a running subtotal under $60”

Consumer reviews and confirms

human_review

“the consumer reviews and confirms before it lands in the cart”

Eval harness clusters failures

feedback_loop

“background agents cluster failures, perform deep-dive investigation, and generate reports for the team. Some reports identify bugs in the assistant itself”

Reported outcome

DoorDash Assistant is rolling out to select U.S.
areas on iOS; patterns from early consumer exposure show around seven in ten messages are discovery requests and most sessions are multi-turn; AI-assisted development roughly tripled weekly pull-request volume in the final pre-launch weeks.

Reported metrics

Share of messages that are discovery requestsaround seven in ten messages are some form of discovery

weekly PR volume increase — early sprintsdoubled

weekly PR volume increase — final pre-launch weeksroughly tripled

Reported stack

Model Context Protocol (MCP)Google's Agent Development Kit (ADK)Vercel AI SDK

Source

https://careersatdoordash.com/blog/building-doordash-assistant-an-engineering-overview/

Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

DoorDash Assistant is rolling out to select U.S.

What tools did this team use?

Model Context Protocol (MCP), Google's Agent Development Kit (ADK), Vercel AI SDK.

What results were reported?

Share of messages that are discovery requests: around seven in ten messages are some form of discovery; weekly PR volume increase — early sprints: doubled; weekly PR volume increase — final pre-launch weeks: roughly tripled (source-reported, not independently verified).

What failed first in this deployment?

How is this ecommerce ops AI workflow structured?

Consumer submits request → Consumer memory retrieval → Live catalog and store search → LLM orchestration assembles response → Grounding validation via tool call → Widget rendered to consumer → Consumer reviews and confirms → Eval harness clusters failures.

DoorDash Assistant is rolling out to select U.S.areas on iOS; patterns from early consumer exposure show around seven in ten messages are discovery requests and most sessions are multi-turn; AI-assisted development roughly tripled weekly pull-request volume in the final pre-launch weeks.