Agentic AI
Google Debuts Agent Compute Engine to Run AI Agents in Isolated, Credential-Free Sandboxes
Google's new managed service executes AI agents in isolated microVMs with security controls that keep API keys and tokens separate from agent execution environments.
InfoQ
Google has unveiled Agent Compute Engine, a managed platform designed to run AI agents safely within isolated microVMs while maintaining strict credential separation. The service uses checkpoint and restore technology to pause and resume agent execution, enabling efficient resource use and low-latency recovery of idle workloads.
Security Through Isolation and Separation#
The core innovation centers on keeping sensitive authentication material—API keys, OAuth tokens, and credentials—completely outside agent sandboxes. Instead of storing secrets within the execution environment, Google routes all authenticated requests through an external Agent Gateway, which injects the necessary credentials only when agents need them. This design prevents compromised or misbehaving agents from ever gaining direct access to sensitive data.
The architecture enforces a clear boundary between the orchestration layer (called "the brain") and the sandbox environment (the "hands"). As Google's distinguished engineers explain, the harness runs entirely outside the sandbox, ensuring that untrusted guest code can only execute commands explicitly dispatched by the harness and stream results back. This separation also creates a verifiable audit trail linking each tool invocation to the harness-initiated request that triggered it.
Technical Foundation: MicroVMs and Copy-on-Write#
Agent Compute Engine uses copy-on-write memory sharing to allow multiple agents to run on a single OS template, reducing memory overhead. Idle agents can be paused via fast checkpoint-and-restore snapshots, minimizing both the cost and latency of resuming work later. This approach tackles three key challenges at scale: ensuring sandboxes remain free of exploitable credentials even when agents fail, reducing the resource and speed penalties of managing idle workloads, and preserving persistent workspaces across sessions and OS updates.
Governance and Observability#
Google has also expanded its agent governance toolkit with centralized controls. The stack includes:
- Agent Gateway: traffic control and policy enforcement
- Agent Registry: tracking and approval of agents and their available tools
- SPIFFE-based Agent Identity: cryptographic authentication and audit logging
- : monitoring of model calls and tool use

