# Cloud Run adds singleton instances for long-lived, individually addressable workloads Google has added a fourth Cloud Run resource type: a manually managed singleton container with its own stable URL, continuous runtime and no autoscaling, aimed at personal agents and other workloads that do not fit request-driven serverless services. Cloud Run instances sit between autoscaling serverless services and a small VM. They run one individually addressable container continuously, can be stopped and restarted, and use shared CPU economics; Google’s launch example prices 1 vCPU plus 1 GiB running for 30 days at $5.70. - Status: Active - Published: 2026-08-31T06:26:36+12:00 - Updated: 2026-08-31T06:26:36+12:00 - Categories: Artificial Intelligence, Cloud & Infrastructure, Hosting, AI Agents, Cloud Platforms, Compute & AI Infrastructure - Tags: AI agents, Cloud Run, Google Cloud, serverless, singleton compute - Canonical HTML: https://beyondthe.news/dossiers/cloud-run-singleton-instances-long-lived-agents ## What changed Google made Cloud Run instances available in Preview on August 25 and published a detailed launch explanation on August 27. An instance is a top-level singleton Cloud Run resource: it runs exactly one container with no autoscaling, has an individually addressable HTTPS URL, and is managed through explicit create, stop, start, update and delete lifecycle operations. Google says instances can run continuously for up to seven days before an automatic restart and can then continue indefinitely through restart policy. Its launch example puts a continuously running 1 vCPU, 1 GiB instance at $5.70 for 30 days. ## Why it matters Cloud Run previously optimized around request-driven services, run-to-completion jobs and worker pools. The new resource fills a different shape: one long-lived process that needs a durable network identity but does not justify a traditional VM or horizontal autoscaling. That is especially relevant to personal AI agents, lightweight control planes and low-duty-cycle workers. The trade-off is that this is not persistent compute in the storage sense: the container filesystem and in-memory state are disposable, so durable state still belongs in external storage. ## Cloud Run now has a managed singleton resource Cloud Run instances are distinct from ordinary service instances. They are top-level resources that are created and operated individually rather than produced by an autoscaler. Each has its own stable `run.app` URL and can be started, stopped, restarted, updated or deleted directly. Google describes the resource as a lightweight serverless VM for workloads that need one stable runtime rather than a horizontally scaled service. ## The resource is shaped around long-lived agents Google’s launch material explicitly uses personal agents such as OpenClaw and Hermes as examples. These systems often need one process available for one user, with bursts of work separated by idle periods. A normal Cloud Run service wants to scale to zero or scale out with requests; a VM stays allocated and brings more operating-system management. Cloud Run instances instead keep one managed container alive and give it a stable endpoint. ## The low monthly example comes from shared CPU behavior Google says a 1 vCPU, 1 GiB instance running continuously for 30 days costs $5.70 in its launch example. The resource model uses shared CPU: workloads can burst to their allocation, while Google documents a 6.25% baseline outside the burst budget. That makes the shape attractive for agents that are mostly waiting and occasionally spike, but compute-heavy continuous workloads need their own benchmark rather than assuming the headline example applies. ## Long-lived does not mean durable local storage An instance can run for days or weeks through automatic restarts, but updates restart the container and Cloud Run instances do not have persistent disk storage. Stopping an instance also destroys in-memory files and unpersisted state. Builders therefore need Cloud Storage, a database, NFS or another external persistence layer for anything that must survive lifecycle events. ## Key details - Cloud Run instances entered Preview on August 25, 2026. - An instance is a standalone singleton Cloud Run resource with no autoscaling. - Each instance can receive its own stable HTTPS `run.app` URL. - Instances are managed individually through create, stop, start, restart, update and delete operations. - Google’s launch post says one instance can run continuously for up to seven days, with automatic restart configured by default. - The resource model uses shared CPU with a documented 6.25% baseline outside burst allocation. - Google’s example prices 1 vCPU plus 1 GiB running continuously for 30 days at $5.70. - Cloud Run instances do not provide persistent disk; stop or update operations can destroy unpersisted local and in-memory state. - The feature remains subject to Google Cloud Pre-GA terms. ## Builder takeaways - Use an instance when the workload genuinely wants exactly one long-lived process and a stable endpoint; use services, jobs or worker pools when scaling or execution semantics fit those resources better. - Treat the $5.70 figure as a specific Google example, not a universal workload price; benchmark CPU-heavy agents against instance limits and alternative VM/serverless options. - Persist agent memory, files, queues and checkpoints outside the container because local state does not survive every lifecycle event. - Test automatic restart behavior and reconnection logic before relying on an instance for unattended agents. - Restrict ingress and IAM on the stable per-instance URL rather than assuming a singleton endpoint should be publicly reachable. ## What to watch - General availability and any change to runtime limits or support guarantees. - Whether Google adds persistent-volume options or stronger stateful primitives for instances. - How shared-CPU burst behavior performs for real agent workloads with irregular tool use. - Pricing comparisons with small Compute Engine VMs, Cloud Run worker pools and third-party agent-hosting products. - Whether managed singleton instances become a broader pattern across serverless providers. ## Uncertainties - The feature is Preview, so API shape, limits and support guarantees can change. - Google’s $5.70 example reflects one continuous 1 vCPU/1 GiB configuration and does not establish cost for compute-intensive agents or external services they call. - The stable URL and restart policy improve continuity, but applications still need explicit durable storage and recovery logic across restarts. ## Sources - [Cloud Run release notes](https://docs.cloud.google.com/run/docs/release-notes) — Google Cloud · primary · 2026-09-01T00:00:00+12:00. Primary release announcement for system-managed Agent Identities and automatic Agent Registry registration. - [Deploy personal AI agents with Cloud Run instances](https://cloud.google.com/blog/products/serverless/introducing-cloud-run-instances) — Google Cloud · primary · 2026-08-27T00:00:00+12:00. Launch explanation covering one-instance semantics, seven-day continuous runtime, stable URLs, stop/resume behavior and the $5.70 example. - [Deployment options and resource model](https://docs.cloud.google.com/run/docs/resource-model) — Google Cloud · primary. Official comparison of services, jobs, worker pools and instances, including singleton addressing, restart policies and shared CPU behavior. - [Create and manage Cloud Run instances](https://docs.cloud.google.com/run/docs/instances/create-and-manage-instances) — Google Cloud · primary. Lifecycle documentation confirming manual management and the lack of persistent disk storage.