# Google details parking idle agent sessions to cut GKE node costs

published: 2026-11-16T17:06:02.279Z
topic: infrastructure
source: https://developers.googleblog.com/agents-are-sessions-not-requests-size-gke-node-pools-for-ai-agent-sessions/
url: https://aiagentnews.io/stories/google-details-parking-idle-agent-sessions-to-cut-gke-node-costs

Google Developers Blog says coding agent sessions spend most of their time waiting for prompts, so one Pod per session wastes node capacity on GKE. It describes Agent Substrate and AX, which suspend idle sandboxes to object storage and free their machines, and gives a cost model showing when that saves money.

Google Developers Blog argues that an agent session stays open all day but works only part of it. Citing a study of GitHub Copilot agent traces, it says agents wait for the user about 80% of the time. Under its assumption of 2 prompts an hour over an 8 hour day, the agent works 22% of the day, and its cost model puts 100 coding agents with one Pod each at about USD 1,288 a month on 34 c3-standard-4 nodes.

Agent Substrate 1.0, now generally available on GKE, can snapshot an idle actor to object storage and free its worker, and the next request restores it on any warm worker. AX sits above it and manages Tasks and Workspaces, and its ax suspend command keeps the workspace while processes start fresh on resume. Agent Substrate does not suspend idle actors by itself, so the harness or the service in front of it must call the suspend API after an idle timeout. When every worker is held, a request waits up to 5 seconds by default before failing with HTTP 503.

In the cost model, a 2 minute idle timeout with a 15 node warm pool and one always-on node for the control plane saves 40% of the node bill at 2 prompts an hour. The saving falls as prompt rates rise, to 4% at 4 prompts an hour, and one Pod per session becomes cheaper at around 4.5 prompts an hour. The blog notes that CPU overcommit changes the comparison and that its cost model does not price it.

## Why it matters if you build agents

If you run coding agents on GKE, measure prompts per hour before choosing between one Pod per session and a parked warm pool, since the cost model's break-even is near 4.5 prompts an hour. Parking needs your own code to call the suspend API, and the default pool autoscaling follows worker Pods rather than suspended actors, so you may need a scheduled job to scale the pool down. Compare Pods and actors at the same CPU guarantee, because overcommit can make the cheaper option differ.
