Google’s AlloyDB preview provisions isolated, serverless PostgreSQL instances for agents in seconds, providing read-only production access and scaling back to zero.

Google introduced PostgreSQL for agents in AlloyDB in preview, with sandboxed database instances provisioned in seconds. Google describes an architecture engineered to support thousands of serverless instances with up-to-the-second, read-only access to production data, separated from primary, standby, and production read-replica instances.

The instances access live database data through a unified storage layer in Colossus, Google’s exabyte-scale distributed storage system. Google reports sub-millisecond input/output (I/O), terabit-per-second aggregated scan throughput, and support for more than 3 million queries per second; the supplied excerpt does not establish these as per-instance figures or provide benchmark conditions.

Each instance uses the full AlloyDB PostgreSQL engine, including indexes, SQL capabilities, and vector, full-text, and spatial search. Google also describes access to BigQuery and Spark for lakehouse analytics without complex extract, transform, and load (ETL) pipelines, alongside identity and access management (IAM) authentication, Virtual Private Cloud (VPC) Service Controls, customer-managed encryption, and auditing.

When agent tasks finish, the agent instances scale back to zero, with billing described as covering active reasoning loops rather than continuously running agent instances. A companion architecture article frames physical isolation, sub-second data freshness, and sub-millisecond storage I/O as design requirements; its example of 1,000 compute nodes demanded within seconds is a workload scenario, not a separately demonstrated deployment result.