Google turns the infrastructure behind Gemini’s agents into a cloud service

Google is set to offer its infrastructure powering Gemini agents as a cloud service, announced at the Linux Foundation’s AGNTCon+MCPCon in San Jose. This new offering, Agent Compute Engine, will provide AI agents with hardware-isolated microVMs that combine the speed of containers with the security of a sandbox. Competitors such as E2B, Daytona, and AWS also provide agent sandboxes, but Google’s service is particularly noteworthy due to its scale and focus on security.

Google’s Agent Compute Engine has been in use internally for Gemini, Antigravity, and model training pipelines, and will soon be available to Google Cloud users in preview later this year. The company’s engineers note that agentic workloads have quickly become one of the fastest-growing consumers of compute and storage across Google’s fleet, necessitating a secure, efficient, and scalable solution.

Agent Compute Engine essentially gives every agent a persistent, sandboxed computer, allowing them to run shell scripts, headless browsers, and code with root permissions. It is designed for large-scale use cases like eval loops, AI benchmarking, and reinforcement learning, and also enables the creation of sandboxed backends for Google’s own services.

Google emphasizes the security of Agent Compute Engine, ensuring that agents cannot break out of their sandboxes and that all traffic is end-to-end encrypted. For users who prefer not to use Google’s managed service, Agent Substrate, an open-source runtime for GKE, is available.