Google puts its AI agents in isolated microVMs
Google launches Agent Compute Engine, a service giving each AI agent its own isolated microVM. The engine, already used for Gemini, will be available in preview later this year with an open Kubernetes version.
On October 22, 2026, Google unveiled Agent Compute Engine, an infrastructure service that assigns each AI agent its own hardware-isolated microVM. Initially designed for Gemini, then opened to Google Cloud customers in preview later this year, it targets a key issue in agentic AI: running autonomous assistants without compromising security or costs. On the same day, Google also detailed Agent Substrate, an open-source runtime for Kubernetes designed to handle spikes of micro-tasks that traditional clusters struggle with.
Chaque agent avec sa propre boîte fermée, sans le syndrome du tiroir à secrets
Agent Compute Engine is based on a simple idea: an agent doesn’t need a whole server, but it does need an environment that doesn’t overflow. Google therefore provides it with a microVM, that is, a lightweight virtual machine, with root access, dedicated kernel, and default isolation. When the agent awaits a model or tool call, the machine pauses, then resumes later with its state preserved. This approach combines the speed of containers with the isolation of a VM, without mixing the two to the point of creating a shaky attack surface.
In its technical analysis, Google explains that it designed this infrastructure after observing that traditional VMs remain too slow for agents that start and stop continuously, while containers are insufficient when full control of the kernel and file system is required. The service relies on Google’s Titanium security stack, Cloud Hypervisor, and a nearly stateless host, with a read-only filesystem and cryptographic verification of hardware and firmware at startup. Access to external services goes through a proxy outside the sandbox, which adds credentials only when authorized. The message is clear: it assumes an attacker controls everything inside, then gives them as little as possible. Not glamorous, but effective.
Une réponse à l’explosion des agents, pas juste à un besoin de confort
This launch matters because agent workloads quickly become costly at scale. Google says these uses are already among the most dynamic consumers of compute and storage in its infrastructure. To limit costs, the company relies on automatic pausing of inactive VMs and a form of shared memory between instances to reduce the cost per agent. Google’s cloud version must also adapt to very different cases: code assistants, evaluation loops for labs, benchmarking, or Google’s internal workloads.
Agent Substrate extends the same logic on the Kubernetes side. This open-source runtime adds a lightweight control plane to manage millions of short-duration tool calls, with snapshotting of an inactive agent’s memory and returning the machine to the pool. Google claims it can offer up to ten times the density of traditional compute and that a test with 200,000 agents sustained 500 activations per second, with suspension and resumption in under 500 milliseconds. Teams can choose microVM or gVisor, while cryptographic identity and egress proxies maintain control over access. The whole system integrates with existing environments, including Cloud Run, GKE, and managed endpoints.
Google n’est pas seul, mais il vise plus large que le simple bac à sable
The market didn’t wait for Google to invent disposable agent environments. E2B, Daytona, or AWS with Bedrock AgentCore already offer sandboxes, often based on Firecracker or similar variants. The difference here lies in integration: Google turns an internal component that already runs Gemini, Antigravity, and its own training pipelines into a cloud product. Meanwhile, Agent Substrate targets teams that want to keep their Kubernetes infrastructure and avoid rewriting all access control.
Another interesting detail: Google adds oversight tools such as token tracking, model routing, and support for third-party identities like Ping Identity and SailPoint. The issue is no longer just about “running” agents, but knowing who controls them, with what identity, and over what scope. This is where the promise becomes concrete for companies: fewer unnecessary exposures, more governance, and a bit less cold sweat when an agent decides to read an email in the wrong place. Integrators dislike surprises; CISOs even more so.
Key points
- Google announces Agent Compute Engine on October 22, 2026.
- Each agent receives a hardware-isolated microVM.
- Agent Substrate promises up to 10 times more density.
- Google test: 200,000 agents, 500 activations/second.
- Claude Code and Hermes are compatible with Substrate.
- Ping Identity and SailPoint are supported.
By the numbers
- 10 times — density announced by Google for Agent Substrate.
- 200,000 agents — benchmark cited by Google on October 22, 2026.
- 500 milliseconds — average resumption time announced for agents.
- 14 days — maximum duration of AWS AgentCore EC2 sessions mentioned for comparison.
- 8 hours — ceiling for standard AgentCore sessions cited by Google.