Google Cloud quietly pushed a set of releases that, taken together, make persistent agents and singleton runtimes first-class citizens — and they also add a new billing and operational surface teams must account for.
Agent Substrate on GKE is now available for evaluation and non-production use, with production support gated behind a limited allowlist GA. That distinction matters: you can try the substrate runtime on test clusters today, but if you want it driving production workflows you’ll need to be on Google’s radar. Equally important and easy to miss in the release notes: Agent Platform session and memory‑bank compute are metered under a new Agent Platform compute SKU. In short, these agents are not free background threads — they are a billable compute footprint that sits alongside your nodes and serverless invocations.
Cloud Run also moves closer to hosting long‑lived workloads. The release notes mark Cloud Run Instances as a Preview capability targeted at long‑lived, individually addressable workloads — the serverless model is expanding from short request/response functions to dedicated singleton runtimes that can hold state and stay addressable. If you read the Cloud Run Instances Preview as “serverless processes that behave like tiny VMs,” you’re not far off; this is explicitly intended to support persistent agent‑style workloads. (We wrote about the implications of singleton runtimes earlier: Cloud Run Instances (Preview): Dedicated singleton runtimes for long‑lived AI agents.)
Finally, the GKE Extended channel now defaults new cluster creation to version 1.35.6. That’s an operational footnote with teeth: teams creating clusters in the Extended channel will get that release by default, and if you plan to evaluate Agent Substrate on GKE you'll want your control plane and node images aligned with that stream.
Metering session and memory‑bank compute under an Agent Platform compute SKU means persistent agents show up in cost reports differently than ephemeral Cloud Run invocations or batch jobs. Teams who treat agents like low‑priority background tasks will be surprised when they see sustained Agent Platform compute charges. It’s straightforward metering, but the mental model is new: agents are persistent compute, not cheap callbacks.
Security and lifecycle follow naturally. Allowlist GA is the right call: Google should keep production rollouts conservative while they work out quota, isolation, and platform hardening. But the allowlist approach also means larger organizations with procurement or security needs will have to plan for a staged adoption path. If your security team hasn’t already discussed process isolation, credential scoping, and ephemeral keying for long‑lived agents, start that conversation now.
Operational implications you should act on now:
- Revisit cost models: include Agent Platform compute SKU usage in monthly forecasts and attribute persistent agent hours to teams or features. Don't assume serverless economics.
- Treat Cloud Run Instances like addressable nodes: add monitoring, liveness, and restart policies appropriate for long‑lived services rather than short functions.
- If you want production Agent Substrate access, prepare for the allowlist: operational readiness, security reviews, and a contact plan with Google.
This is overdue and inevitable. The cloud ecosystem has been drifting toward an “agent‑first” architecture — from managed agent services to platform automation — and GCP is making that explicit: runtimes for agents, billing for their persistent footprint, and serverless primitives that let them stay addressable. That clarity is useful. What isn’t useful is ambiguity in pricing and quotas; if your team doesn't treat agents as first‑class billable infrastructure now, they will be a surprise line item later.
Expect the next wave to be about controls: per‑agent quotas, stronger network isolation, and finer‑grained audit and lifecycle APIs. If you’re a platform team, stop thinking of agents as scripts and start treating them like services — with capacity planning, security boundaries, and billing owners. Google provided the primitives this week; it’s now on teams to treat them with the operational discipline they require.
For a bit more context on where Google is taking these changes, see our earlier coverage of Agent Substrate and Cloud Run Instances in GKE 1.36: GKE 1.36: Agent Substrate on GKE and Cloud Run Instances for Persistent Agents.