Google Agent Compute Engine brings Gemini's agent sandboxes to Cloud

Google will make the infrastructure that powers its Gemini agents available to Google Cloud customers. Agent Compute Engine, announced at the Linux Foundation's AGNTCon+MCPCon in San Jose, is due to enter preview later this year, as The New Stack reports.

The service gives every AI agent a hardware-isolated microVM that Google says combines the speed of containers with the security of a sandbox. It began as internal infrastructure for Gemini, Antigravity and Google's own model-training pipelines. Google says agentic workloads have become one of the fastest-growing consumers of compute and storage across its fleet.

Security-first design

In a technical deep dive, Google distinguished engineers Rohit Jnagal and Ramanathan Venkatapathy argue that traditional VMs boot and resume too slowly for agents, while containers lack the isolation and root-level control some agents need. Agent Compute Engine gives each agent a persistent sandboxed computer where it can run shell scripts, use a headless browser and write code, in loops that may run for weeks. Once a tool call finishes, the microVM is checkpointed and uses no active resources until the next call.

Each session runs in a single-tenant, memory-safe microVM with its own guest kernel. The environment holds no secrets in memory or on disk, and its only identity is short-lived and session-scoped. The microVMs have no standard network interface. Traffic goes through a proxy outside the sandbox, where Google's Agent Gateway checks the user's delegation and the sandbox's identity before adding credentials. The design rests on Cloud Hypervisor and Google's Titan and Titanium IPU security platform. The team's rule: assume an attacker controls everything inside the sandbox, and leave nothing there to use.

A crowded field

Sandboxes for agents are not new. E2B and Daytona offer them, and AWS provides Bedrock AgentCore. Google itself already offers a managed Agent Sandbox on GKE and sandboxes within its Gemini Enterprise Agent Platform. Beyond coding agents, Agent Compute Engine targets large-scale uses such as model evaluation loops, benchmarking and reinforcement learning.

Teams that prefer Kubernetes can use Agent Substrate, an open source GKE runtime first announced in May. It builds on Agent Sandbox and adds a lightweight control plane. Google claims up to 10 times the density of traditional compute, and in a 200,000-agent benchmark it reports 500 activations per second, with agents suspending and resuming in under half a second.