Building Ask DoorDash (Part 4): A Platform for Building and Evolving Agents - DoorDash

This post details the platform built for creating and evolving domain-specific agents within Ask DoorDash. It describes the architectural decision to centralize shared capabilities like orchestration, memory, model access, tracing, evaluation infrastructure, and rollout controls, while domain teams own instructions, skills, and model choices. Key components include a Gateway for request translation and streaming, an Orchestrator for routing turns to specialized domain agents and pinning conversations, and a skill system to modularize instructions and tools, reducing context token costs. The Model Context Protocol (MCP) layer provides safe and reusable access to DoorDash services. The platform enables faster feature addition (e.g., Reservations agent in one week) and quicker model improvements (e.g., p50 turn latency cut by 35% and then another 40%).