Google Cloud has released Agent Substrate, an open-source runtime built to run agent sandboxes by the millions, and made it generally available on Google Kubernetes Engine. The pitch is specific: Google says the system delivers 10 times the container density of standard runtimes, with resume operations completing in under 500 milliseconds and more than 500 suspend and resume activations handled per second. Those are Google’s own benchmarks, published in a company blog post from engineers Alex Zakonov and Tim Hockin.
A laptop running Claude Code or Codex faces a completely different problem than a fleet of hundreds of thousands of agents that write code, invoke tools, and execute for hours without a human in the loop, which is the gap Google says most coding-agent products have not had to confront yet. Untrusted code running without kernel isolation can escape its host and steal credentials, according to Google’s post, and reinforcement-learning rollouts or benchmark runs can each spawn thousands of sandboxes within a minute, more traffic than a general-purpose scheduler is built to absorb.
Its answer is to strip machine management away from agent execution and let each sandbox run on its own terms. Two isolation options are on offer: hardware-backed Cloud Hypervisor for full microVM boundaries, or gVisor for a lighter-weight kernel boundary, and either sits behind a gateway that keeps credentials out of the agent’s reach as it handles egress traffic. When an agent goes idle, Agent Substrate snapshots its state to disk and Cloud Storage in a few hundred milliseconds, packing more than 1,000 dormant agents onto a single host, according to Google. Node recovery, autoscaling, and cluster reliability stay with Kubernetes, while a purpose-built data plane takes over the sub-second suspend and resume traffic itself.
Nous Research, which builds the Hermes Agent, is an early adopter. Google credits OpenRouter’s usage data with putting Hermes first among AI agents overall, first in productivity, first in coding, first in CLI tools, and first for personal-agent use, a ranking sourced to OpenRouter rather than any independent audit. “Our experience with Agent Substrate gives us confidence the architecture can scale efficiently as agent workloads grow,” said Hervé Bizira, Nous Research’s chief business officer, in a quote Google supplied in the same post.
Google is also pairing the runtime with two of its own products. Axion, its Arm-based processor line, claims up to 30 percent better cost efficiency than rival clouds on sandbox workloads, and Filestore agent volumes, a shared-storage option, attaches in milliseconds and locks files so multiple agents can write without colliding. Google is the source for both figures, not an outside benchmark.
The runtime itself is open source and runs on any Kubernetes cluster, not just GKE, which matters more than the density numbers do. Google is not asking developers to trust a black box: teams can inspect the isolation model and the suspend and resume logic themselves before deciding whether the density claims hold up on their own workloads. General availability on GKE is limited to an allowlist for production traffic, with broader non-production access open now, so the real test of those numbers arrives once teams outside Nous Research start running their own agent fleets against it.
Google Cloud, “Agent Substrate available on GKE,” published on the Google Cloud blog, September 17, 2026.
