Google just sketched the runtime stack that will change how platform teams run agents: an open-source GKE Agent Substrate that promises 10x the density of normal container runtimes, sub-500 ms resume latency, and the ability to support millions of sandboxes — paired with Cloud Run instances (Preview) for singletons that can run continuously up to seven days, and Gemini Pro and lightweight Gemini variants available across Vertex AI and the Gemini API surfaces.
The impressive-sounding metrics matter because they change trade-offs. Google says the substrate can handle hundreds of suspend/resume activations per second and sub-500 ms resume times, which makes it practical to keep agent state paused on dense nodes and bring agents to life on demand. That’s a different optimization from typical Kubernetes workloads: instead of autoscaling pods up and down with minutes of cold-start pain, you keep ultra-cheap, paused sandboxes and resume them in fractions of a second. Google describes kernel- and network-level isolation for these sandboxes — the implicit promise is you can run many untrusted agent instances on shared nodes without reintroducing legacy host-level risk.
Cloud Run instances (Preview) fills the other side of the spectrum: addressable, dedicated singleton runtimes with stable HTTPS URLs, no autoscaling, automatic restarts, and up to seven days of continuous runtime. Google also published an example price for a shared-vCPU configuration as a point of comparison; treat published examples as illustrative, not as a commitment to a new billing model across GCP. Still, it’s the unit economists will use when mapping agent patterns to dollars — short-lived, dense sandboxes for bursty work plus a handful of cheap singletons for persistent coordination looks like the shape of things to come.
On the model side, Gemini Pro and lighter Gemini variants are available in preview across Vertex AI and the Gemini API. Put the pieces together and you get a clear reference architecture: GKE Agent Substrate for massively parallel, isolated execution; Cloud Run instances for addressable, persistent runtime agents; and managed Gemini models for reasoning and state updates.
This is not just packaging — it forces platform teams to think about new responsibilities. If you run agent sandboxes at the density Google is promising, you need nontrivial scheduler and packing logic, node image and kernel hardening, and observability that scales to millions of suspend/resume events per minute. Telemetry pipelines that work for long-lived services won’t cut it: you’ll want event-first traces and compact state snapshots. And yes, IAM and network policy must become per-agent primitives: the attack surface moves from service-to-service to agent-to-resource.
Two practical consequences you’ll face immediately:
- Identity and secrets: short-lived sandboxes + sub-second resumes mean your credential distribution system must handle ephemeral secrets and rotate aggressively. Treat agent identity like workload identity, not a service account you hand out once.
- Cost governance: the Cloud Run instance example shows persistent singletons can be inexpensive individually, but at scale they add up. If your control plane or coordinator agents leak or duplicate, run-rate cost will surprise you faster than CPU hours do.
Opinion: Google’s move here is the right call. The industry has been cobbling agent runtimes out of generic containers and serverless primitives for two years; formalizing a substrate for tens of millions of sandboxes is overdue. However, the productization also accelerates a painful truth: platform teams that don’t invest in agent-level controls — identity, egress, telemetry cardinality, and cost observability — will be handed a new, high-velocity attack surface and bill shock.
If you want to read more about the pieces, I covered the GKE sandbox direction and the Cloud Run instances preview in separate posts: GKE Agent Sandbox for AI agents: isolated, stateful runtimes on GKE and Cloud Run instances (Preview): dedicated singleton runtimes that run up to seven days.
Prediction: within 12 months we’ll see agent-level billing APIs, per-agent network egress quotas, and managed sidecars for telemetry/secrets that are installed automatically by the substrate. If you own platform engineering, start treating execution as a first-class networked service — not just another pod.