Google announced Agent Compute Engine on October 22, 2026, in a Google Cloud blog post. The service is purpose-built compute for AI agents, already powers Gemini's agent sessions, and will reach enterprise customers in preview in the coming months. On the same day, Google open-sourced the Agent Substrate runtime for Kubernetes and extended its agent governance capabilities to more environments.
Agents work differently, so they need a different computer
Ordinary programs run and finish. Agents write code, execute it, drive browsers and call tools over many rounds, pausing while the model thinks or a human replies. Traditional virtual machines isolate strongly but take seconds or minutes to start, pause and resume; containers start fast but share the host kernel, which is a shaky home for untrusted agent code.
Agent Compute Engine gives every agent a hardware-isolated microVM: its own guest kernel and root access, a hardware-isolated hypervisor boundary, and Google's Titanium offload chip underneath for defense in depth. Each microVM runs with short-lived credentials and a cryptographically verified identity, and its workspace persists on its own Hyperdisk across sessions.
Density comes from pausing and sharing memory
Cost is the second storyline. When an agent enters its thinking phase, the microVM signals idleness and the host pauses it with a fast checkpoint, restoring it from a snapshot when needed instead of burning an idle machine. Agents launched from the same template share memory copy-on-write, taking private copies only of what they change. The open-source Agent Substrate brings the same design to Kubernetes: Google reports up to 10x density over traditional compute, and in a benchmark of 200,000 agents it sustained 500 activations per second with suspend and resume latency under 500 milliseconds, working alongside frameworks such as Claude Code, OpenClaw and Hermes.
Governance additions include token and cost controls in Agent Gateway, policy-based model routing, and integrations with identity providers such as Ping Identity and SailPoint, extending from cloud runtimes to cross-cloud, on-premises and endpoint coding agents. For teams moving agents from pilots to production fleets, the message is direct: competition is shifting from the models themselves to whoever can run thousands of agents safely and cheaply.
