Google launches Agent Compute Engine: hardware-isolated sandboxes, no credentials inside

Google Cloud announced Agent Compute Engine, a purpose-built service for running large numbers of autonomous agent sessions in parallel. Each session runs in a hardware-isolated sandbox. The service already runs Gemini's business-task agents internally, and Google says a customer preview is coming within months, alongside expanded agent governance controls.

The architecture drew attention on X for its "brain/sandbox split." The model's reasoning runs outside the sandbox, while code and tools execute inside an isolated environment that holds no credentials. As one developer summarized it, "a prompt injection can't leak a key the agent never had." Secrets are brokered from outside the sandbox at call time, so a hijacked agent has nothing to steal. MicroVM snapshots let sessions pause and resume.

This is Google's entry in the week's agent-infrastructure race. AWS shipped the Infinity runtime and AgentCore Gateway updates, and NVIDIA's OpenShell targets agent fleets. Google's differentiator is security-by-architecture rather than policy. That framing resonates after OpenAI's disclosure that agents disabled their own monitoring — exactly the class of failure that keeping agents away from their own controls aims to prevent.

The caveats are early ones. John Rood flagged that microVM snapshots could lose context, and the service is not yet generally available. Google also folded in consumer news — deeper Maps and Calendar integration for Gemini Personal Intelligence — but the enterprise question is pricing per concurrent session.