At AGNTCon + MCPCon, Google introduced Agent Compute Engine, a compute service purpose-built for AI agents. It already powers Gemini as an agent for work tasks and will be available in preview in the coming months. The company also extended Gemini for business agent governance to cover agents running across Google Cloud, cross-cloud, on-prem, and endpoint environments. The post frames the shift from single-turn prompting to interactive agents as an infrastructure problem: running autonomous agents securely at planet scale requires a new execution and control paradigm.

Agent Compute Engine was built for Google’s heavy agentic workloads, where agents write and execute code, run tools, drive browsers, orchestrate enterprise workflows, and iterate rapidly. It gives every Gemini agent its own high-performance, secure computer. Key design points: microVMs boot quickly, copy-on-write memory sharing lets many agents run from one operating system template, and fast checkpoint/restore pauses idle microVMs. When an agent finishes a tool call and enters the thinking phase, the microVM signals idle, moves from active to idle state, and can later resume from snapshots. On security, the service runs with Titanium, Google’s offload infrastructure processing unit, and provides hardware-isolated Cloud Hypervisor boundaries with a guest kernel and root access. Each microVM uses the agent’s own identity with short-lived credentials, has a cryptographically verified ID, and gets service credentials injected outside the microVM. Integration with Agent Gateway adds observability, policy, and security evaluation. For persistence, each microVM has its own Hyperdisk, and enterprise connectivity is governed by the same Google Cloud identity model for unified identity and auditability. Gautham Thambidorai, VP and Google Fellow for Gemini, says these microVMs provide the responsiveness, hardware isolation, and workload density needed for interactive agent sandboxes at scale.

For platform teams that want Kubernetes-native customization and portability, Google also announced Agent Substrate on Google Kubernetes Engine, an open-source execution runtime with a zero-trust architecture that separates agent execution from machine management. When an agent is idle, Agent Substrate snapshots its active memory, returns the machine to the pool, and resumes the agent when needed. Reported highlights include up to 10x higher compute density compared with traditional compute while keeping full kernel isolation and identity-based network security, and a benchmark of 200,000 agents sustaining 500 activations per second with less than 500 ms suspend/resume latency. Each sandbox uses microVM or gVisor isolation, every agent gets its own cryptographic identity, and egress proxies enforce network policy and inject credentials outside the agent. It runs on any Kubernetes cluster and works with any agent framework or harness, including Claude Code, OpenClaw, and Hermes, and it integrates with Filestore agent volumes for shared multi-agent workspaces. Nous Research’s chief business officer, Hervé Bizira, says Hermes Enterprise uses Agent Substrate for per-agent isolation and access control while preserving compute resources.

On governance, the post describes the triple challenge of securing dynamic non-deterministic agent interactions, maintaining default-deny least-privilege access across runtimes, and auditing asynchronous agent actions. Existing Gemini platform governance components include Agent Gateway, Agent Registry, Agent Identity using the SPIFFE standard, and OpenTelemetry-compliant agent observability. The expansion now covers Google Cloud runtimes such as Agent Runtime, Cloud Run, Agent Sandboxes, GKE, and Agent Compute Engine; multicloud and hybrid environments through Envoy-based policy enforcement points; and managed-device endpoints for coding agents like Claude Code and Antigravity CLI via OS-level sandboxing. New Agent Gateway features coming in preview this quarter include token and cost governance with spend attribution and budget guardrails, model routing policies, Service Extensions for datapath programmability, a security partner ecosystem with Ping Identity, SailPoint, and Saviynt, and Express Link for Cloud Run to establish secure agent-to-tool connectivity. L’Oréal’s AI lead engineer, Alexandre Frizzo, notes that Google Cloud’s agent governance gives the company a foundation to scale agentic AI while unifying identity, policy, and observability across runtimes.

The post closes by positioning the combination of Agent Compute Engine, Agent Substrate, and expanded governance as a resilient foundation for protected, high-density AI agents, with preview interest and an open-source repository for Agent Substrate available.