Google Cloud Launches Cloud Run Instances for Long-Lived AI Agents

Google Cloud introduces a new runtime for always-on AI agents

Google Cloud introduced Cloud Run instances on August 27, 2026 as a preview designed for long-lived, stateful workloads such as personal AI agents.

The important distinction is that this is not the normal autoscaling Cloud Run service model. A Cloud Run instance runs as a dedicated singleton runtime with one instance and no autoscaling, making it better suited to agents that need to remain alive between requests.

What Cloud Run instances provide

Google says each Cloud Run instance:

• runs as a single instance with no autoscaling;
• can run continuously for up to seven days, with automatic restart enabled by default;
• receives a stable HTTPS URL that persists across updates and restarts;
• can be stopped when not needed and resumed later.

Google lists a reference price of $5.70 for 30 days for a continuously running instance configured with 1 vCPU and 1 GiB of memory. That price is specific to the configuration and pricing described in Google's launch post, not a universal cost for every agent workload.

Why Google is targeting AI agents

Conventional Cloud Run services scale to zero when requests stop. That is efficient for request-driven web apps, but it can be awkward for a personal agent that expects one process to stay online, preserve working state, and wait for tasks.

Google's launch example uses OpenClaw, an open-source personal AI agent. The company shows it being deployed with a gcloud beta run instances create command, with agent configuration stored in Cloud Storage and credentials supplied through environment variables.

Google also cites long-running agent workloads where demand is bursty: an agent may stay online continuously but only use substantial compute when it receives a task.

Preview status and upcoming SSH access

Cloud Run instances are in preview, not general availability. Production teams should therefore evaluate preview support terms, regional availability, quotas and lifecycle behavior before depending on the feature for critical services.

Google also says SSH access is coming soon for both Cloud Run instances and Cloud Run services. SSH is therefore an announced future capability, not part of the current preview described in the launch post.

Why this matters for agent infrastructure

The rise of always-on AI agents is creating infrastructure patterns that sit between serverless request handling and full virtual machines. Developers often want persistent process state, a fixed endpoint and continuous uptime without managing an operating system or provisioning a VM.

Cloud Run instances are Google's attempt to fill that gap with a managed singleton container runtime.

For teams building personal assistants, coding agents, monitoring agents, messaging bots or other long-running agent processes, the preview is worth testing because it combines a stable endpoint and persistent runtime model with Cloud Run's managed container environment.

The tradeoff is that this runtime intentionally gives up horizontal autoscaling. It is therefore better suited to singleton or low-concurrency agents than large stateless services that need elastic scale.