The app for independent voices
DoorDash is building AI agent swarms that it says will work like ant colonies.
Most companies are still figuring out how to get a single AI agent to reliably complete a task.
DoorDash has already moved through three phases:
Deterministic workflows (simple, predictable, no agentic capabilities)
Single agents (one agent reasons through a problem, queries multiple tools, delivers an answer)
Swarms (peer agents passing work between each other asynchronously)
The swarm model is still experimental. But for a logistics business juggling drivers, restaurants, customers, and real-time pricing, a swarm model could have a significant impact.
What makes all of this possible is their sophisticated AI agent infrastructure (see attached).
Their architecture connects source data (Snowflake, Google Drive, Confluence, BI reports), a vector database for searchable knowledge, a context builder for RAG and few-shot prompting, and a toolkit of specialized tools (SQL, Slack, Wiki, feedback processors).
If you're building agents internally, Doordash is a fine example of why infrastructure is so important.
I go deeper into this and other AI agent use cases in the latest DoP Deep Dive: departmentofproduct.sub…
