DoorDash Details AI Assistant Architecture Built Beyond a Single LLM

DoorDash has published details of the architecture behind Ask DoorDash, its conversational shopping assistant, revealing that the company deliberately moved away from a single large language model approach in favor of a system of multiple specialized agents, persistent consumer memory, and standardized tools. The architecture was described in a series of posts on the company's engineering blog and was later analyzed by InfoQ.

A runtime instead of one model

At the core of the system is what DoorDash calls the Assistant Runtime, a component that coordinates specialized agents responsible for different tasks: catalog search, recommendations, cart building, checkout, and order history. Instead of embedding business logic directly into prompts sent to the language model, the assistant relies on reusable tools built on the Model Context Protocol that connect to DoorDash's existing services.

Separating orchestration from business functions means different AI experiences within the app can share the same integrations while the backend evolves independently. According to reports, the system was built in part on Google's Agent Development Kit, with the MCP layer exposing catalog and transactional functions to all agents at once.

Three layers of memory

The second pillar of the architecture is a consumer memory system made up of three components: long-term memory generated offline from a user's behavioral history, session memory covering the current conversation context, and agentic memory that stores facts explicitly provided by the customer. Data flows into a distributed storage layer with vector search, while the orchestration layer handles retrieving context tailored to the specific task at hand.

DoorDash describes this mechanism as turning fleeting signals from a single session into persistent memory that improves recommendation quality over time. This layer is responsible for the largest measured business impact, as shown in tests run on grocery data.

How they measure quality

The third component is an automated evaluation infrastructure that simulates multi-turn conversations with language-model-generated users and relies on recorded tool fixtures to recreate production conditions without involving real customers. The framework mimics the production environment closely enough to allow independent assessment of orchestration, safeguard mechanisms, and individual domain agents.

Building a useful AI agent is hard. Knowing if it is actually good is even harder - Raghav Saboo, lead of the RecSys and Search team, DoorDash

Thanks to this infrastructure, the company says it cut regression testing time from six hours to twenty minutes and raised its quality score by eight points, while also using the framework to validate model migrations, including one that reduced response latency by 35 percent without sacrificing quality.

What the numbers show

The most concrete data concerns grocery shopping: during a seven-day evaluation, computed consumer memory raised checkout conversion by about 24 percent, increased average basket size by 17 percent, and cut the number of conversation turns required by 7 percent. Cart building was reportedly about five times faster than manually browsing the catalog, and a single well-formed prompt could complete an order in under two minutes. For restaurant search with vague, open-ended queries, the assistant achieved 15 percent higher conversion than standard browsing.

DoorDash publicly launched Ask DoorDash in June 2026, letting customers place orders using photos and free-form descriptions, and the company also integrated the shopping feature directly into ChatGPT. The architecture writeup now published on the engineering blog explains how these features work under the hood and why the company chose a multi-agent approach instead of a single model handling all queries.

Why it matters for the industry

DoorDash's case is relevant for Polish e-commerce and ordering platforms because it illustrates a pattern that large players are increasingly repeating: language models generate the language, but business decisions and access to transactional data stay within specialized, controllable tools. This kind of architecture makes it easier to audit an agent's actions and limits the risk of faulty transactions, which matters when real customer money is on the line.

The publication of these technical details also fits a broader trend in which large consumer platforms openly share the architecture of their agentic systems, partly to attract engineering talent and partly to establish design patterns for the wider industry building AI agents for large-scale production use.