Google Cloud unveiled a computing service designed for running autonomous AI agents at scale, alongside an expansion of its agent governance capabilities across distributed environments.
Dubbed Agent Compute Engine, the hyperscaler's new infrastructure aims to address the performance, security, and resource management challenges associated with running batches of AI agents concurrently.
Much like Meta's Muse agent system, Agent Compute Engine addresses these challenges by rapidly booting microvirtual machines (microVMs), giving individual agents isolated execution environments without requiring dedicated physical machines. This offers an alternative to traditional computing infrastructure which is not optimized for agents which flit between cold starts and a flurry of compute activity: constant code execution, interaction with external tools, navigating browsers, and carrying out multistep workflows.
To improve resource efficiency, Google’s new service uses copy-on-write memory sharing, allowing multiple agents to operate from a common operating system template. It also incorporates checkpointing and restoration capabilities that suspend idle microVMs and release computing resources while AI agents wait for user input, external systems, or model responses.
The microVMs can subsequently resume from VM-based snapshots, reducing infrastructure overhead, and enabling higher workload density. For workloads requiring persistent storage, each microVM can also access its own Google Cloud Hyperdisk, permitting execution environments and agent context to persist across long-running sessions.
On the security side, Agent Compute Engine uses Google's Titanium infrastructure processing unit to provide hardware-backed isolation, while each microVM operates within a Cloud Hypervisor security boundary with its own guest kernel and root access.
Each VM instance also receives a cryptographically verified identity and short-lived credentials, with the latter injected into outbound connections outside the microVM. Google's Agent Gateway tool additionally provides extra monitoring, policy enforcement, and security evaluation, with new services centered on token consumption tracking and request routing.
Broader governance updates for the Google Cloud suite see Agent Gateway bolstered with model routing policies directing requests from users and agents toward different AI models according to organizational requirements.
Outside of Agent Compute Engine and Gateway, Google will introduce lightweight policy enforcement for multicloud and hybrid environments which leverages Envoy, an open-source networking proxy that sits between applications and services.
In a likely response to recent OpenAI discretions, Google's other forthcoming governance capabilities will combine sandboxing at the operating system level with local governance services to extend corporate policies to agents working directly on user machines. Finally, networking tool Express Link for Cloud Run can now allow AI agents to declare the resources they need to access, allowing for authenticated connections and identity tokens to be injected automatically.