Google Cloud unveils Agent Compute Engine and opens Agent Substrate to run enterprise AI agents at planet scale
Built for AGNTCon and MCPCon, the microVM architecture features Titanium IPU isolation, persistent Hyperdisks, and sub-second Kubernetes scheduling tested with 200,000 agents.
Technology news
At AGNTCon and MCPCon North America on Thursday, Google Cloud unveiled Agent Compute Engine, a dedicated execution infrastructure purpose-built to deploy and govern enterprise AI agents at scale. Built directly upon Google’s Titanium offload infrastructure processing units (IPUs), the platform provides silicon-rooted microVM isolation with independent guest kernels and root access, ensuring that non-deterministic, untrusted agent actions cannot compromise host systems or neighboring enterprise workloads.[1]
To resolve the state-persistence bottleneck in long-running agentic tasks, each microVM is backed by Google Cloud Hyperdisks, preserving workspaces and conversational state across multi-step sessions without requiring customer data replication. For organizations standardized on containerized orchestration, Google also open-sourced Agent Substrate for Google Kubernetes Engine (GKE). The zero-trust execution runtime snapshots memory when agents idle and resumes them in under 500 milliseconds, sustaining 500 activations per second across 200,000 tested agents while delivering up to 10x higher compute density than traditional VM allocation.[1]
The open-source Substrate runtime natively supports third-party coding agents and agent harnesses, including Claude Code, OpenClaw, and Hermes Enterprise. Highlighting production validation, Hervé Bizira, Chief Business Officer at Nous Research, noted that Agent Substrate solves multi-tenant agent isolation and extensible access control directly at the platform layer while conserving critical compute resources. The system integrates into an expanded Agent Gateway featuring SPIFFE-based agent identities, OpenTelemetry tracing, and dynamic token cost management.[1]
Our take: The AI industry is hitting a harsh infrastructure reality: running autonomous coding agents inside full persistent virtual machines burns enormous compute when idle, while stateless serverless functions lack the persistence required for multi-step reasoning. By decoupling agent execution from machine management through sub-second memory snapshotting and hardware-backed Titanium microVMs, Google Cloud has delivered the first serious, production-grade cloud operating system for autonomous AI agents. Partnering openly with open-source frameworks like Hermes and Kubernetes gives Google a crucial architectural edge over proprietary walled gardens.