Google Cloud announced a preview of PostgreSQL for agents in AlloyDB, a new architecture aimed at letting AI agents query production databases without destabilizing the operational workloads running on the same system. The stated problem: enterprise databases are mission-critical, but even a few agents running dense reasoning loops can produce unpredictable query surges that overwhelm traditional architectures.

The core mechanism is dynamic provisioning of sandboxed database instances in seconds. These instances get up-to-the-second read-only access to production data, while remaining fully separated from the primary, standby, and read replica instances where production workloads run. A unified storage layer in Colossus, Google’s exabyte-scale distributed storage system, backs these instances. The source claims sub-millisecond I/O, terabit-per-second aggregated scan throughput, and support for over 3 million queries per second. Each instance runs the full AlloyDB PostgreSQL engine, so every index, the full SQL surface, and comprehensive vector, full-text, and spatial search are available to agents.

The pricing model is pay-as-you-go: instances can scale out to absorb sharp spikes of concurrent agent queries and then scale back to zero when agents finish, so customers are billed only for active reasoning loops rather than for around-the-clock dedicated read replicas. The architecture also includes native lakehouse integration without ETL. Agents can run federated queries across BigQuery and Lightning Engine for Apache Spark, joining large lakehouse datasets with up-to-the-second transactional data in AlloyDB.

The announcement positions this against alternative approaches that put page-caching layers above object storage. The source argues those architectures face scaling and cost challenges, can have up to an order of magnitude higher I/O latency, and can be saturated by active multi-agent workloads, leaving production systems vulnerable to agent-induced outages. By contrast, AlloyDB‘s instances remain isolated from production clusters while still accessing fresh operational data, and the full engine supports mixing analytical queries, vector queries, and operational point lookups in the same agent loop.

Security is handled through deep Google Cloud integrations: IAM authentication replaces static passwords, VPC Service Controls isolate traffic, and customer-managed encryption and auditing are supported. The post also includes a quote from Sanjeev Siotia, Executive Vice President and Chief Technology Officer at Manhattan Associates, saying the architecture has been a game changer for their supply chain platform because it lets them deploy collaborating agent networks against inventory and order data with sub-second freshness while keeping core transactional processing untouched.

PostgreSQL for agents in AlloyDB is available now in preview. The post provides links to documentation and a sign-up form, but the source gives no further availability details.