Size a GKE node pool for the sandboxes holding workers, not simply the number of agent sessions left open. Google’s model estimates nearly 40% lower node costs from parking idle sessions at two prompts per hour in a scenario of 100 coding agents, but the result depends on its workload and us-central1 pricing assumptions. Agent Substrate’s retry and placement rules matter when workers are scarce.
Size for the workers your active or not-yet-suspended sandboxes occupy, then compare the busiest hour with your own prompt rate.
Why an open agent session does not always need a worker
A persistent agent session can keep its conversation open between prompts while the sandbox doing the work is idle. A Kubernetes worker, by contrast, is compute capacity: when an actor is running on it, that capacity is in use. Parking an idle actor frees its worker without ending the session.
Google’s guide uses a two-minute idle timeout in its detailed pool example. In that model, sessions hold a worker for nearly 29% of their open time. The pool is sized for the 99th percentile of simultaneously worker-holding sessions, with 15% headroom and an assumption that sessions behave independently. These are inputs to Google’s example, not a fixed recipe for every workload. Google’s GKE sizing guide
What parking preserves—and what scaling to zero changes
In Agent Substrate, an actor is the isolated sandbox for an agent, and a reusable worker is a warm Kubernetes Pod that can host multiple actors. SuspendActor snapshots an idle actor to object storage and releases its worker. The guide says memory is captured by default, with files as the fallback. The agent harness or a service in front of it must call the suspend API after its chosen idle timeout; suspension does not happen automatically.
Scaling a Pod to zero while retaining a persistent volume is a different tradeoff: files can persist, but the process and its memory do not. The next prompt starts a new agent process. Google’s guide also describes AX as preserving /workspace through a volume snapshot while processes start fresh on resume. Google’s AX runner guide
How Agent Substrate handles a full worker pool
Agent Substrate v1.0 enables request parking by default. When a resumed actor encounters a transient resource-exhaustion or control-plane-unavailable error, the router retries within a five-second budget per resume flight. That is a bounded retry budget, not a guaranteed five-second delay before failure. A capacity-related HTTP 503 can still occur if the problem persists through the budget or the default parking lot reaches its limit of 1,024 concurrent parked requests. Agent Substrate’s request-parking documentation
Worker capacity also has to match actor limits. Agent Substrate’s v1.0 API guide requires CPU and memory limits for each actor. Placement needs enough remaining declared worker capacity to cover those limits plus runtime overhead, and actors share the worker’s capacity. Pod requests are not the actor-sizing field. The same guide says GPU devices are temporarily unsupported for actors: placing a worker on a GPU node does not pass a GPU into the actor. Agent Substrate’s API guide
What Google’s model estimates at different prompt rates
Google models 100 coding agents receiving two prompts per hour over an eight-hour workday. In its one-Pod-per-session baseline, each Pod has 1 vCPU and 1 GiB of memory; Google estimates 34 c3-standard-4 nodes and about $1,288 per month at Google Cloud on-demand list prices in us-central1 as of November 2026. The same guide gives these modeled outcomes as prompt rates change:
| Prompt rate | Google’s modeled result |
| 2 prompts per hour | Nearly 40% lower node cost with parking |
| 4 prompts per hour | 4% savings with parking |
| Above about 4 prompts per hour | One Pod per session becomes cheaper |
The node counts in Google’s examples describe different views of the model: one comparison reduces the 100-agent requirement from 34 nodes to 16, while its detailed burst-sized configuration uses 15 warm nodes plus one always-on node for the Agent Substrate control plane and PostgreSQL. Neither count is a general GKE sizing prescription. Google’s GKE sizing guide
To apply the estimates, Google recommends measuring prompt rates from a week of gateway logs and comparing the busiest hour with the model’s approximate crossover near four prompts per hour. Its guide also recommends scaling the worker pool to zero after hours; in the model, running the pool for all 730 hours of a month costs nearly twice as much as the one-Pod-per-session alternative.