Google Cloud Run instances target long-lived AI agents with singleton serverless compute

Google Cloud has introduced Cloud Run instances, a preview offering designed for long-lived, stateful workloads such as personal AI agents. The company announced the feature on August 27, 2026, positioning it between two common but imperfect options: stateless serverless services that scale to zero when requests stop, and dedicated virtual machines that require continuous management and full-time compute spending.

The new product is aimed at a practical infrastructure gap created by AI agents. Traditional Cloud Run services are well suited for request-driven web applications, APIs and jobs that can scale horizontally. Many agents behave differently. Tools such as OpenClaw or Hermes may run continuously for one user, monitor events, remember local configuration, wait for messages and then spike briefly when asked to complete a task. For that pattern, autoscaling to zero can interrupt the assistant, while a virtual machine adds operating-system maintenance, firewall configuration and unnecessary cost.

Cloud Run instances provide a dedicated singleton runtime on Cloud Run. Google says each instance runs as a single copy with no autoscaling, can run continuously for up to seven days with automatic restart policy configured by default, receives a stable HTTPS URL that remains unchanged across updates and restarts, and can be stopped when not needed and resumed later. That gives developers a managed home for a persistent agent without requiring them to operate a full VM.

The pricing example is also central to the announcement. Google says a Cloud Run instance with 1 vCPU and 1 GiB of memory running continuously for 30 days costs 5.70 dollars. The service uses shared vCPU with burst budgets, which fits agents that are alive for long periods but do not constantly consume heavy compute. That distinction is important because many personal or small-team agents spend most of their time waiting for input, webhooks or messages, then use short bursts of model calls, tools or browser actions.

Google’s example deployment uses OpenClaw, an open-source personal AI agent that many users initially run on laptops. The problem with a laptop-based agent is obvious: it stops when the machine sleeps, loses availability when the user travels and cannot easily serve as a stable endpoint for messaging tools. Cloud Run instances let a developer package that agent into a container, attach configuration through Cloud Storage and expose a consistent URL. Google also said SSH access for Cloud Run instances and Cloud Run services is coming soon through private access.

The early customer signal comes from OffDeal, an AI-powered investment bank for small businesses. Google says OffDeal uses Cloud Run instances as primary infrastructure for a long-running agent and reported an 88 percent reduction in cold starts. That is the kind of metric agent builders watch closely. Multi-step agents can make dozens of calls inside one task, and startup delays can compound into a poor user experience even when the model itself is fast.

Cloud Run instances show that AI agents are forcing cloud platforms to rethink serverless assumptions. The old model optimized for stateless request handling and elastic scale. Agent workloads often need continuity, stable identity, background operation and controlled cost for mostly idle runtime. If the preview matures, it could give developers a cleaner deployment path for personal assistants, internal workflow bots and small autonomous services that need to stay awake without becoming full infrastructure projects.