Google Launches Agent Compute Engine for Secure AI Operations

Google has announced the launch of Agent Compute Engine, a managed service that facilitates the operation of AI agents within isolated microVMs. This service includes checkpointing capabilities, enabling the suspension and resumption of execution environments.
The design of Agent Compute Engine prioritizes security by ensuring that credentials, API keys, and OAuth refresh tokens remain outside the agent sandbox. Instead, it utilizes an external Agent Gateway to authenticate requests and inject necessary secrets.
According to Google, the Agent Compute Engine addresses three key requirements at a global scale: maintaining sandboxes free from exploitable credentials, minimizing the cost and latency associated with resuming idle agents, and preserving persistent workspaces across sessions and operating system updates.
The infrastructure team highlighted that the service achieves these goals through copy-on-write memory sharing, allowing multiple agents to operate on a single operating system template. Additionally, it can pause idle microVMs via fast checkpoint and restore mechanisms.
In a detailed overview, Google engineers Ramanathan Venkatapathy and Rohit Jnagal outlined the architecture of the service, which includes microVM isolation, credential management, agent identity, storage, and scheduling at scale. They emphasized the strict separation between the orchestration harness, which runs outside the execution sandbox, and the sandbox itself, which executes commands dispatched by the harness.
This architectural boundary creates a deterministic audit trail that correlates each harness-initiated tool call with the expected compute activity within the sandbox.
The Agent Compute Engine supports four primary workload categories: interactive consumer and enterprise agents that execute code and browse the web; persistent Linux workspaces for autonomous software engineering; large-scale parallel environments for model evaluation and reinforcement learning; and lightweight, sandboxed execution environments for dynamic logic in Google services.
Venkatapathy and Jnagal also noted that Google employs this platform internally for various applications, including Gemini, Gemini for Business, and the agentic development platform Antigravity.
In addition to Agent Compute Engine, Google has introduced Agent Substrate, an open-source, Kubernetes-native runtime for running AI agents in isolated environments. Agent Substrate manages the lifecycles of sandboxes independently from the underlying compute infrastructure, allowing idle agents to be suspended while preserving their state for future execution.
For further insights, Harel Asaf, an AI specialist, discusses the implications of the Agent Compute Engine in his article on the topic.
Source: InfoQ