How DoorDash Built an AI Shopping Assistant That Doesn't Rely on the LLM Alone

DoorDash's Ask DoorDash is not a simple Q&A system relying solely on an LLM, but an architecture where multiple specialized agents work together. The core is an "Assistant Runtime" that coordinates agents from different domains and exposes business functions through a shared MCP (Model Context Protocol) layer. These functions include catalog search, recommendations, cart, checkout, order history, and user memory, all backed by DoorDash's existing backend services.

The advantage of this design is that business logic is not embedded in prompts but invoked through reusable tools. Multiple AI experiences can share the same integration layer, while backend capabilities can evolve independently. DoorDash emphasizes that agents not only need access to user data but also "need the right context at the right time for the right task."

To provide precise context, DoorDash introduces an intelligence layer managing personalization through three memory systems: long-term memory (generated offline from historical behavior capturing preferences like favorite cuisines and dietary restrictions), session memory (maintaining conversational context), and agentic memory (storing facts explicitly stated by the user).

Relevant memories are retrieved via semantic vector search, ranked, and injected into prompts. This design separates memory management from model reasoning, allowing each interaction to receive tailored information. DoorDash reports that in a 7-day evaluation with computed user memory, cart checkout conversion improved by about 24%, basket size increased by 17%, and conversation turns decreased by 7%.

DoorDash recognizes that "building useful AI is hard, but judging whether it is truly good is even harder." To address this, they built an automated evaluation framework using LLM-generated users and recorded tool fixtures to simulate stateful customer conversations. The framework mirrors the production runtime and can independently evaluate orchestration, guardrails, and domain agents.

Currently, the platform runs over 2,000 automated evaluations daily, improving quality scores by 8 points and reducing regression testing from 6 hours to 20 minutes. Additionally, the framework validated a model migration that reduced latency by 35% without sacrificing quality. This provides quality assurance for rapid iteration.

Besides the conversion improvements from memory, the assistant achieves 15% higher conversion for open-ended restaurant queries. Co-founder Andy Fang states that Ask DoorDash builds carts roughly 5 times faster than manual operation; users can complete a cart with a single prompt in under 2 minutes.

DoorDash also optimizes latency and reliability through deterministic actions (e.g., not invoking LLM for versioned artifact updates) and confirmation workflows. The overall architecture follows their engineering model: domain teams build specialized agents, platform teams maintain orchestration, MCP tools, memory, evaluation, and shared components. Early results demonstrate the effectiveness of this decoupled architecture, but long-term performance and cross-domain generalization remain to be seen.