Google has officially launched Agent Compute Engine, a new Google Cloud service designed to handle the massive compute demands of agentic AI. The platform provides every AI agent with its own hardware-isolated microVM, achieving container-like speed while maintaining the security boundaries of a virtual machine. This move addresses the critical infrastructure gap that emerged as AI shifted from simple question-answering to autonomous action-taking.

The Problem with Traditional Compute

The Gemini team discovered that traditional VMs were too slow for agents that constantly start, pause, and resume, while standard containers lacked the kernel-level isolation required for secure code execution. Agents spend most of their time waiting for model responses or user input, making the reservation of full CPU and memory resources for every idle session prohibitively expensive. Google needed a system that could cheaply maintain the working state of millions of agents while waking them up instantly when needed.

Zero-Secrets Architecture

Security is the cornerstone of Agent Compute Engine, built on a defense-in-depth model that assumes an attacker controls everything inside the sandbox. The architecture ensures that no persistent secrets, OAuth refresh tokens, or cloud service account keys ever reside in the microVM's memory or on disk. Instead, the system uses an ephemeral, session-scoped SPIFFE certificate, while long-lived user credentials are dynamically swapped by the Agent Gateway only after validating the session's identity and egress policies.

Application-Aware Efficiency

To maximize density, Google co-designed the sandbox command runtime with the host executor to create native application awareness. The infrastructure detects when an agent enters the 'thinking' phase and immediately transitions it to an idle state via a fast checkpoint-and-restore model, consuming zero active resources. This allows a single physical node to host orders of magnitude more living agent microVMs than conventional VMs, supported by Direct Access (DAX) and Copy-on-Write memory clones that deduplicate static pages across the host.

Key Takeaways

  • Agent Compute Engine provides hardware-isolated microVMs with container-speed boot times for AI agents.
  • The 'zero-secrets' architecture ensures no credentials exist inside the sandbox, mitigating risks from malicious prompts.
  • Application-aware idle detection allows massive scaling by preserving agent state without active resource consumption.
  • Integration with Google Workspace and Cloud allows agents to access data without moving it or opening firewall ports.

The Bottom Line

Google's Agent Compute Engine finally solves the 'last mile' problem of agentic AI: infrastructure that is as secure as a VM but as efficient as a container. This isn't just a product update; it's the architectural foundation required to make autonomous agents viable at planet-scale.