Google Cloud announced Cloud Run instances on August 27, 2026, adding an option for durable, stateful workloads, including AI agents. The service is aimed at workloads that need a persistent execution model rather than a short-lived process.
The practical change is continuity. Google Cloud says Cloud Run instances runs a single instance without autoscaling, and that continuous execution can last up to seven days. For an AI agent handling work over an extended period, that creates a defined runtime boundary without introducing automatic scaling into the service itself.
A stable endpoint with a clear limit
Each Cloud Run instance receives an HTTPS URL that remains unchanged after updates and restarts. Clients can therefore continue to target the same address when the instance lifecycle changes, which gives stateful workloads a predictable endpoint.
That stable address does not remove the need to handle restarts. An automatic restart policy is configured by default, so the process still has a restart event to account for even though its URL does not change. The announcement describes a consistent endpoint, not uninterrupted execution.
The single-instance design is the main trade-off. It can give a durable agent a straightforward execution shape, but capacity will not expand automatically through autoscaling within this model. Workloads that need more than one instance must address that requirement in their own architecture.
The cost example is narrowly defined
Google Cloud gives one concrete cost reference in the announcement: an instance with 1 vCPU and 1 GiB of memory, running continuously for 30 days, costs $5.70. That figure applies to the stated resource profile and duration, so it should be treated as a baseline for that configuration rather than a universal price for every agent deployment.
For teams considering long-running AI agents on Google Cloud, Cloud Run instances brings together four specific properties: a single instance, up to seven days of continuous runtime, automatic restarts and a persistent HTTPS URL. Those properties address durability and endpoint continuity, while the absence of autoscaling remains the boundary that shapes how the service can be used.
Cloud Run instances is therefore a focused infrastructure option for stateful workloads, not a general answer to every deployment pattern. Its usefulness depends on whether a workload fits the seven-day runtime and single-instance model, with the published $5.70 example offering a concrete starting point for the smallest configuration described.