Google Gives Every AI Agent Its Own microVM: 200,000 Agents, 500 Wake-Ups a Second

Google Cloud used the AGNTCon and MCPCon stage this week to introduce Agent Compute Engine, a compute service built for one workload in particular: agents that start, think, act and go idle thousands of times an hour. The pitch is that agents need a different kind of computer than either virtual machines or containers provide.

The isolation model is the headline. Every agent runs inside its own hardware-isolated microVM with a dedicated guest kernel and root access, sitting behind a hypervisor boundary, with Google Titanium offload silicon underneath. Each microVM runs with short-lived credentials and a cryptographically verified identity rather than long-lived secrets baked into the sandbox, and its workspace persists on its own Hyperdisk across sessions, so an agent asked on Tuesday about a spreadsheet it analyzed on Monday does not start from scratch.

The economics come from two tricks. First, application-aware idle detection: because Google co-designed the sandbox command runtime with the host, the infrastructure knows when an agent has finished a tool call and is waiting on the model, then checkpoints it to a snapshot and reclaims the machine instead of leaving it spinning. Second, memory deduplication: hosts map pre-warmed snapshot images directly into physical memory and hand out copy-on-write clones, so hundreds of agents running the same base image, Python runtime and browser share the same physical pages and pay only for the deltas they modify.

Google puts the combined effect at up to 10x the density of conventional compute. In a benchmark of 200,000 agents, the company says the platform sustained 500 activations per second with suspend and resume latency under 500 milliseconds. Those are vendor figures and have not been independently reproduced, but the mechanism behind them is legible: if idle agents cost close to nothing, the cost of a large fleet stops being a function of how many agents exist and starts being a function of how many are thinking at once.

Google also open-sourced Agent Substrate, the runtime that brings the same design to Kubernetes, and says the system already works alongside agent frameworks including Claude Code, OpenClaw and Hermes. Governance features moved in the same release: token and cost controls in Agent Gateway, policy-based model routing, and integrations with identity providers such as Ping Identity and SailPoint, extended from cloud runtimes to cross-cloud, on-premises and endpoint coding agents.

The service is not speculative. Google says it already runs its own Gemini agents on it, in both consumer and business configurations, and Clement Farabet, a vice president at Google DeepMind, tied the design to Antigravity, describing instant snapshot-and-fork primitives as what makes long-horizon software engineering feel effortless. Enterprise customers get access through a preview that Google says is coming in the coming months, without a fixed date.

The strategic reading is straightforward. As agents move from demos to fleets, the constraint shifts from which model answers best to who can run tens of thousands of them safely and cheaply, and the vendors selling that substrate are positioning themselves as the layer that outlives any single model generation. Google is also, implicitly, arguing that containers alone are the wrong primitive for untrusted agent code, a point that lands differently for every competitor whose agent sandbox is a shared kernel and a promise.