Databases Are Being Redesigned for AI Workloads

Postgres handled it fine. Until the agent started issuing thousands of embedding lookups a second.

Every assumption underneath a normal database is a human assumption. Queries arrive at human rates, from a known set of applications, with someone available to notice when the answers look wrong.

Agent workloads break all three. That’s what’s driving the current wave of database redesign, and 3 changes matter for anyone choosing a store this year:

1. Vector search moving into the general-purpose engines. Postgres with pgvector, and equivalents elsewhere, handle a lot of retrieval without a separate system. A dedicated vector database earns its place at genuine scale or with heavy filtering needs, and rarely before that.
2. Separation of storage and compute becoming the norm, so a burst of agent traffic doesn’t compete with your nightly transformations for the same resources.
3. Metadata and semantics as first-class citizens. Agents need column meaning, units and caveats readable at query time. Your catalog becomes an interface.

For most SMEs the honest move this year is boring: keep the engine you have, add the vector extension, put the semantics in a catalog. Rewrites earn their cost only when the access pattern proves them necessary.

What in your stack would break first if an agent queried it at machine speed?