What happened
Google Cloud introduced Agent Compute Engine, an infrastructure service designed to host artificial-intelligence agents in isolated microVM sandboxes. The sandboxes run on stateless appliances that keep customer files only in an immutable file system. Firmware integrity is verified with Google’s custom Titan chips, and the processors include mitigations against side-channel attacks. Network traffic must pass through an isolated IPC boundary and multiple proxies before reaching the data centre, according to engineers Ramanathan Venkatapathy and Rohit Jnagal.
The service detects when a sandbox goes idle and powers it down, then quickly reactivates it when needed. This allows a single physical node to host many more live agent microVMs than conventional virtual machines. Features such as cloning, snapshot/forking, rewind and rollback are built in to avoid resource duplication.
Agent credentials are stored outside the sandbox. When an agent requires login, a separate Google Cloud service called Agent Gateway retrieves the credentials and supplies them on the agent’s behalf, preventing hackers from stealing sensitive data even if they compromise a sandbox.
Google plans to roll out Agent Compute Engine in the coming months, alongside a new version of Agent Gateway that adds model-routing and token-management tools to control inference costs.
Why it matters
The microVM approach gives the efficiency of containers while retaining the strong isolation of virtual machines. By shutting down idle sandboxes, organisations can reduce hardware utilisation and associated power consumption, addressing both cost and sustainability concerns. Storing credentials outside the sandbox and enforcing strict network guardrails reduces the attack surface for AI agents, a growing vector for cyber-threats.
The integration of cloning and snapshot capabilities means developers can spin up many agent instances without duplicating underlying resources, speeding up AI-driven automation workloads. Token-management in Agent Gateway adds financial governance, helping enterprises keep AI inference spend under control.
What this means for enterprise IT teams
IT teams can deploy more AI agents per server without sacrificing security, thanks to the microVM isolation and idle-shutdown logic. The immutable appliance firmware and Titan-chip verification simplify compliance audits, while the external proxy model provides clear points for traffic monitoring. Credential handling via Agent Gateway removes the need for agents to store secrets locally, easing secret-management processes.
The upcoming model-routing feature lets teams direct each request to the most suitable AI model, improving latency and reducing unnecessary compute. Token-management tools give a single source of truth for spend tracking, aiding budgeting and cost-optimisation efforts.
In the words of Google Cloud engineers Ramanathan Venkatapathy and Rohit Jnagal: "All inbound and outbound traffic must traverse an isolated, point-to-point IPC boundary."
The bottom line
Google Cloud’s Agent Compute Engine offers a secure, efficient platform for AI agents, cutting power use and hardware costs while strengthening protection against malware and credential theft. The service’s built-in lifecycle and cost-control features give enterprise IT departments the tools they need to scale AI workloads responsibly.
Sources
- siliconangle.com