Every autonomous Gemini assistant running in Google's cloud will soon get its own slice of hardware-level isolation. Google has unveiled Agent Compute Engine, an infrastructure service designed to lock down individual AI agents inside isolated micro virtual machines.
The system already underpins Gemini behind the scenes. In the coming months, Google plans to open the engine for public preview, letting outside developers deploy automated software workers without letting them mingle in shared memory spaces.
Running autonomous agents reliably has created an expensive headache for cloud providers. When software agents execute code, browse websites, or handle sensitive customer records, isolating them usually requires dedicated virtual machines. Doing that at scale eats massive server capacity and racks up steep infrastructure bills.
Google claims it has worked around that bottleneck by leaning on its custom Titanium chips. Each microVM draws hardware isolation directly from Titanium silicon, keeping processes strictly partitioned from neighboring tasks.
Efficiency gains come from how these virtual environments rest. Whenever an agent pauses between actions, its microVM suspends automatically, relinquishing machine capacity to other workloads. Shared memory across inactive instances helps shave running expenses further. In an internal trial involving 200,000 distinct agents, Google reported that suspending and resuming instances took under 500 milliseconds.
Packaging Agent Substrate for Kubernetes
To spread the architecture beyond its proprietary setup, Google is releasing an open-source Kubernetes runtime dubbed Agent Substrate. According to the company, the runtime allows servers to pack ten times as many autonomous agents onto a single physical machine compared to traditional virtualisation setups.
Packing denser clusters of software workers is only half the battle. Enterprises also want tighter reins on what those automated tools access once they are turned loose on internal databases. To address governance worries, Google is layering in built-in tracking for token expenditure and dynamic model routing.
The service also connects directly to existing corporate access frameworks, adding support for external identity platforms including SailPoint and Ping Identity. Google confirmed these management safeguards will span Cloud Run, Google Kubernetes Engine, and managed cloud endpoints as the engine rolls out toward wider availability.
ADFiled by The AI Desk
Models, assistants and the companies and chips behind them, reported from what was released and what was claimed, with the difference kept clear.
More from this desk →
Be the first to comment
Join the argument. No password, just your email or a passkey.