Inside Agent Compute Engine: Building the compute foundation for the agentic era
Google introduces Agent Compute Engine, a new cloud service designed to support AI agents with isolated, persistent microVMs, addressing scalability and security challenges in agentic AI workloads.
Google identifies the rise of agentic AI as a foundational shift requiring new compute infrastructure, noting that traditional VMs and containers are ill-suited for the rapid, persistent, and isolated environments agents demand. The company highlights that agent workloads now consume significant compute and storage resources, prompting the need for a dedicated platform to handle their unique requirements at scale.
The new service, Agent Compute Engine, provides each AI agent with a hardware-isolated microVM that combines the speed of containers with the isolation of VMs. This design allows agents to maintain persistent state, resume quickly, and operate securely without exposing credentials or sensitive data within the sandbox environment.
Google outlines three primary use cases for the platform: interactive consumer and enterprise agent surfaces, autonomous software engineering workspaces, and large-scale evaluations for reinforcement learning. The service also supports lightweight sandboxing for internal Google applications, enabling safe execution of dynamic logic.
The platform addresses security concerns by leveraging Google’s Titan silicon and Titanium IPU for hardware-rooted trust, ensuring cryptographic isolation and integrity. Each agentic session runs in a dedicated microVM with zero persistent secrets, mitigating risks from untrusted or adversarial code while maintaining high performance and scalability.