DoorDash cut latency in its AI-powered grocery reorder flow from about 16 seconds to 2 seconds, according to InfoQ's report on the company's architecture. The change also doubled shopping list views and increased the flow's order rate by approximately 34%, InfoQ reports. The fix was not a faster model — it was moving the expensive reasoning out of the request path entirely.
The mechanism is precomputation. Instead of rebuilding a recurring-purchase shopping list from scratch on every request, DoorDash now processes 90 days of purchase history alongside consumer profile signals offline, using batch LLM inference to identify likely recurring purchases. Deterministic validation checks those generated results before they're stored as reusable shopping list data. By the time a consumer opens the reorder flow, the heavy lifting has already happened.
What runs live is narrower: a serving path that retrieves the prepared list and resolves each item against the current product catalog. InfoQ describes a fallback chain — the system first tries to match the original store item, then a catalog product identifier, and finally a search term — which lets it handle discontinued or changed listings without re-running historical reasoning on every call. This is the core trade: the system still invokes a Grocery agent when a request needs genuine reasoning, such as finding cheaper alternatives based on current prices, promotions, package sizes, inventory, and consumer preferences, but simpler edits — changing quantities, removing a product — update the shopping list directly without another LLM round trip.
The reorder flow sits on top of what InfoQ calls DoorDash's shared agent platform, which provides orchestration, memory, model access, tracing, evaluation infrastructure, and rollout controls, while domain teams keep ownership of agent instructions, skills, tools, and evaluation criteria for their own slice of the system. DoorDash reports that adding its Reservations agent on top of this shared platform took one week, which the company says is approximately ten times faster than building its initial domain agents from scratch — a number about reusability, not about inference cost or latency.
The architecture also treats state differently than a typical chat-based agent loop. A DoorDash engineering post cited by InfoQ states that "a consumer's direct edit is authoritative: the next agent turn starts from the edited list, not from what was said earlier." The shopping list is a structured artifact both the consumer and the agent can modify directly, rather than a conversation history the agent has to re-parse to infer current state — which is part of why simple edits don't need to round-trip through an LLM at all.
Twisha Jain, a DoorDash engineer, framed the underlying design call in a LinkedIn post quoted by InfoQ: "Better agentic systems aren't always about more reasoning; they're about deciding what belongs in the agent loop and what can happen ahead of time." That's the actual architectural decision being reported here, and it's also the limit of what's disclosed — InfoQ's piece gives no model name, no token counts, no inference cost figures, and no detail on how large the batch jobs are or how often the precomputed lists are refreshed. The 16-to-2-second latency figure, the 34% order-rate lift, the 90-day history window, and the one-week, roughly-ten-times-faster Reservations rollout are the numbers InfoQ discloses; the underlying serving stack's cost profile is not among them.
The piece also notes, without elaborating, that Instacart's AI shopping assistant handles meal planning, list creation, and product discovery using similar personalization signals — a reminder that the interesting part of DoorDash's case isn't that an agent touches groceries, it's that most of the agent's work was moved out of the live request entirely.
The takeaway for an architect reading this as an orchestration pattern: before you reach for a faster model or a bigger context window to cut agent latency, check how much of the loop is actually re-deriving something that was true yesterday — and move that part offline.