The agentic era of Postgres databases

Agents are turning database creation into a machine behaviour. The next opportunity is making safe, realistic copies of production data routine for agent development.

Every database product of the last 15 years has fit into one of three eras, defined by who creates the database and how long it lives. Databases go back decades to the time of the DBA managing big metal mainframe computers, but our story begins in the early-modern-cloud era of 2009. This is when AWS launched RDS as their first managed database service.

RDS Postgres didn't follow until November 2013, but the model had been set with the operational burden of backups, patching, and failover, all handled by the cloud provider (GCP and Azure eventually built equivalents). Despite this revolution in operations, the unit of provisioning remained the same - you still had to choose an instance size, paid for it regardless of usage, and expected it to live for years. One company may have a handful of database instances, and creating one was a real event.

This era is still with us and it's still where most production data lives, but the new "database as an API" era started around 2021. This was when Neon launched the next generation of database providers. Neon pioneered stateless Postgres compute on top of a distributed storage layer, so an instance could be created in milliseconds and scaled to zero when idle. Billing moved to real usage (sometimes billed as serverless).

Branching also became possible, due to the copy-on-write storage architecture, but whilst it was possible to create database copies for every pull request, the problem of dealing with production data remained. The real revolution was therefore more about the operational model of serverless scaling and billing (again).

PlanetScale took a different route to the same place. It built its reputation on Vitess, the open source system that shards MySQL at YouTube scale. But developers had moved to Postgres, and being MySQL-only became a constraint: in March 2024 PlanetScale cut staff and retired its free tier to focus on paying customers, and hobbyist mindshare drained to Neon, Turso and Supabase. PlanetScale has since added Postgres support, differentiating itself with the performance that comes with bare metal. Did we come full circle?

Supabase also belongs in this era by distribution rather than architecture. Underneath, each project is still a dedicated, provisioned instance - era 1 in a better interface - and its branching is schema-only by default. But it became the default backend for the app-builder generation.

But the Supabase model of rapid instance creation set the stage for what has become the focus of the new era of Postgres database: agents. The change in this is not the architecture - it is the customer. Databases are now created by software operating at rapid speed. When Databricks acquired Neon in May 2025 for a reported ~$1B, it said more than 80% of Neon databases were created automatically by AI agents - up from 0.1% two years earlier. A single customer, Replit, now creates 50,000–70,000 Neon databases a day. Supabase says it adds over 4 million databases a month, around 70% of them created by agents or AI tools. Database creation is now a machine behaviour.

Two technical responses have emerged.

The first is to make the database so small it can be created without thinking. Cloudflare's D1 (SQLite) allows 50,000 databases per account on paid plans, with per-user and per-tenant databases as the intended design.

PGlite goes further: the whole of Postgres compiled to WebAssembly, under 3 MB, running inside a browser or an agent sandbox with no server at all. Its weekly downloads grew from 1M to 13M in twelve months, and in August 2026 Databricks acquired its maker, Electric, explicitly to embed Postgres inside every agent sandbox.

The second response is still mostly missing - agents doing development work, but needing realistic production data. It cannot have production data itself - that is where the personal information lives - so the new workflow needs to be: clone production, anonymize the sensitive fields, hand the clone to the agent, and throw it away afterwards.

This is exactly what Xata was doing before it was acquired by ClickHouse. They were in the market for a long time, so I suspect they made the right technical bet, but just got the timing wrong. Perhaps they'll be able to develop this as part of ClickHouse Managed Postgres, which went GA in November 2026. PGRun is a new example of this type of architecture, but Postgres-native from the beginning.

The era of the DBA is long gone, but classic cloud instances still have immense data gravity. The serverless separation of storage and compute tried to dislodge that approach, succeeding in more modern applications, but still not the majority. The AI era has much more momentum and has a better chance of pushing the model forward. Whoever makes it routine to hand an agent a safe copy of production - anonymised well enough to satisfy a regulator, realistic enough to catch real bugs - will become a critical part of the agentic database era.