Google details Agent Compute Engine for custom agents

Google said it launched Agent Compute Engine at AGNTCon + MCPCon and described it as a purpose-built compute service that already powers Gemini as an agent for work tasks.[1] The company says customers will soon be able to use the service to power their own custom agents, with preview availability expected in the coming months.[1] Google frames the need around agents acting in digital environments, which can require filesystems, language runtimes, browsers and network access.[2]

Google says each AI agent receives its own hardware-isolated microVM, designed to operate at container speed.[2] The company says microVMs boot rapidly so agent loops and evaluation sweeps do not have to wait on infrastructure.[1] Copy-on-write memory sharing lets multiple agents run on one operating-system template, according to Google.[1] When an agent finishes a tool call and enters its thinking phase, the microVM signals that it is idle; Agent Compute Engine can then pause it through fast checkpoint and restore.[1]

Google says the isolation boundary is a hardware-isolated Cloud Hypervisor, with a guest kernel and root access inside each microVM.[1] Each microVM runs its agent’s identity using short-lived credentials and receives its own cryptographically verified ID.[1] Google says service credentials are injected outside the microVM at egress, rather than held inside the sandbox.[1][2] Each microVM also has its own Hyperdisk to persist execution workspaces and agent context.[1] Agent Gateway integration is intended to support observability, policy and security evaluation.[1]

Google argues that traditional compute primitives are not designed to run hundreds of millions of agents at the required cost and latency.[2] It says traditional virtual machines can take seconds to minutes to start, suspend and resume, while containers are fast and lightweight because they share the host operating-system kernel.[2] Google says coding agents and developer assistants can need complete isolation as well as kernel- and root-level control.[2] The company also says it is extending Gemini for business agent-governance capabilities to cover agents wherever they operate.[1]