Database Provisioning Is Now an AI Agent's Job

Originally published on DevToolHub.

At Neon's 2024 GA, 30% of new databases were created by AI agents. By the May 2025 acquisition it was over 80%, per Databricks — the company that bought Neon for roughly $1 billion partly because of this shift.

The Numbers

97% of database branches on the platform now come from natural-language agent requests, not human CLI commands. Multi-agent workflow usage grew 327% in a five-month window in 2025.

Lakebase's Speed

Databricks rebuilt Neon as Lakebase: Paxos-based "safekeepers" for writes, "page servers" for fast reads. Branches provision under 500ms and new instances under a second, handling 12 million launches a day.

Tiger Data's Agentic Postgres

A separate approach: zero-copy forks on "Fluid Storage," pgvectorscale for vector search, and a new pg_textsearch extension for BM25 keyword search — plus an MCP server so agents provision databases from plain-language prompts.

The Limits

pgvector is CPU-only, and memory becomes the ceiling in the tens of millions of vectors. Branching was built for a handful of dev branches, not hundreds of ephemeral agent branches per minute. If agents provision their own branches, cap the count and lifetime per agent — retry loops create branches faster than any human would.

Why It Matters Beyond Neon

SQL Server 2025's native VECTOR type, Oracle AI Database 26ai's built-in vector search, and pgvector on RDS/Cloud SQL all point the same direction: vector search and database provisioning are converging into general-purpose relational databases.

Full writeup: devtoolhub.com/database-provisioning-ai-agents