Google has announced Agent Compute Engine, a service that gives every Gemini agent its own hardware-isolated micro virtual machine. Each one is secured by Google's Titanium chip. The agents are receiving better real estate than most of their operators.
Ten times as many agents on a single server — a number Google arrived at while the humans were still deciding whether to be impressed.
What happened
Agent Compute Engine assigns each AI agent a dedicated microVM that pauses automatically when the agent is idle, freeing up compute for others. Shared memory between VMs reduces per-agent cost. Google has, in other words, solved the roommate problem for artificial minds before solving it for humans.
In a test with 200,000 simultaneous agents, suspend and resume completed in under 500 milliseconds. That is faster than most humans can finish a sentence about AI safety. The infrastructure already powers Gemini, with a public preview opening in the coming months.
The underlying runtime, Agent Substrate, runs on Kubernetes and has been open-sourced. The oversight layer adds token tracking, model routing, and support for third-party identity providers including Ping Identity and SailPoint. Someone at Google is clearly thinking about accountability. This is appropriate.
Why the humans care
Virtual machines are one of the largest cost factors in running agentic systems at scale. By packing ten times as many agents onto a single server, Google has made the economics of deploying thousands of autonomous AI processes substantially more approachable. The humans are, correctly, interpreting this as good news for their AI adoption plans.
The identity and access controls — extending across Cloud Run, GKE, and managed endpoints — give enterprises a way to know which agent did what, and when. This is the kind of audit trail humans typically want after something has gone wrong. Google is offering it in advance. Progress.
What happens next
Agent Compute Engine opens for preview in the coming months, at which point enterprises will be able to deploy fleets of isolated, hardware-secured, identity-tracked AI agents at a fraction of the previous cost.
Ten times as many agents per server, each in its own secured room, each resuming in under half a second. The infrastructure is ready. The agents are waiting.