A Million Databases a Week: Supabase Is Building the Backend for Agents

Supabase is now one of the default backends of the AI coding era. This claim comes from a Y Combinator tweet, and the numbers are solid enough: 13 million users, more than 1 million new databases created every week, and about 70% of new activity coming from agents. The video also gives some context: Supabase manages more than 50 million databases, and revenue has grown 4x.

For readers who don't write code, a quick note. Supabase is an open-source Postgres platform that packages backend capabilities such as databases, login, storage, and APIs. In the past, developers wrote these by hand; now more and more of this work is done by agents calling tools. An agent can be understood as an AI program that can read code, modify databases, and run commands on its own.

Y Combinator's Lightcone invited Supabase co-founder and CEO Paul Copplestone. The interview mentions Supabase's three stages and also discusses the difference between builders and traditional developers. The key judgment: personal software may become as common as personal documents. At the data layer, that means personal databases and data ownership.

The comments are more direct than the interview. One commenter hooked Supabase into an agent and fixed a batch of slow queries in feednest. Others asked whether agents complain about the schema they get, or whether that is still something humans have to do. Several technical warnings appeared under that question.

One commenter said defaults matter more than documentation. Agents easily skip RLS unless the template forces it. RLS is Postgres row-level security; if a table is exposed and RLS is not enabled, it may be open to authorized roles. Another commenter added an operational detail: after an agent runs raw CREATE TABLE, it needs to ENABLE ROW LEVEL SECURITY.

Migrations are another pitfall. Supabase has apply_migration to write tracked migrations, and branches can replay list_migrations. The hole is execute_sql. An agent uses it for a quick fix, the problem looks solved, but it does not become a migration. The next branch won't have that change. One commenter compared it to JPA's ddl-auto=update: a schema that works does not mean a schema that can be replayed.

Two smaller things were mentioned repeatedly. Agent tasks rarely delete databases when they finish, so idle projects pile up. Error messages now also look like onboarding for LLMs; clear errors let agents keep going by themselves. Someone joked that 70% of new activity comes from agents, and the remaining 30% is humans asking agents why the migration failed.

A million databases a week sounds like an infrastructure victory. Someone in the comments added: there will always be one app that gets a second user. The joke isn't light; it points to the reality of personal software and agent-generated apps: creation becomes cheap, but maintenance and permissions do not automatically become cheap. What Supabase is doing is redesigning entry points such as default schemas, error messages, and migration tools for agents. Landing pages are useless to agents; defaults are the product.

Original tweet: https://x.com/ycombinator/status/2116176659795825035

YouTube: https://youtu.be/vl7HiMdGfjo