Google Cloud announced Agent Compute Engine (ACE) at AGNTCon + MCPCon, introducing a compute architecture tailored specifically for AI agents. The service provisions single-tenant microVMs using Cloud Hypervisor to isolate agent execution environments. ACE decouples agent logic into an external harness (‘brain’) and an isolated sandbox (‘hands’) that executes commands, code, and browser sessions without access to direct credentials or secrets.
To optimize efficiency, ACE utilizes memory sharing via copy-on-write and automatically checkpoints idle microVMs to zero active resource consumption when waiting for user or model input. An open-source, Kubernetes-native counterpart named Agent Substrate has been available on GKE since mid-September. Customer previews for Agent Compute Engine on Google Cloud are scheduled in the coming months.
Why it matters
Protects enterprise applications from prompt injection vulnerabilities by isolating sandbox agent execution environments.
Lowers infrastructure costs for agent workloads using microVM idle checkpointing and memory sharing.
Establishes dedicated runtime standards for deploying agentic AI workloads into enterprise cloud environments.
Source: marktechpost.com



