Google Cloud Adds Dedicated Compute and Governance for AI Agents
Google Cloud’s agent strategy now spans both execution and control: Agent Compute Engine supplies dedicated, isolated environments for agents, while Gemini’s governance tools are expanding across cloud, hybrid and endpoint deployments. The shift reflects a practical challenge for production teams: agents repeatedly run tools, pause while waiting, resume work and carry state across sessions. Infrastructure built for steady, conventional workloads can struggle with that pattern.
Announced at AGNTCon + MCPCon, Agent Compute Engine is a purpose-built service that already powers Gemini agents for work tasks. Google says it will enter preview in the coming months. Alongside it, the company is extending Gemini for business governance to cover agents wherever they run. The two efforts address different parts of the same operating problem: providing agents with usable compute while keeping their identities, access and activity under organizational control.
Compute designed around the agent loop
Agent Compute Engine gives each Gemini agent a high-performance runtime built around micro virtual machines, or microVMs. The design targets workloads that write and execute code, use tools, operate browsers and coordinate enterprise workflows. Those tasks can involve frequent changes between active execution and waiting for a person or external service.
That stop-start behavior shapes the service’s efficiency approach. MicroVMs are designed to boot quickly, and copy-on-write memory sharing lets multiple agents use a common operating system template. When an agent finishes a tool call and enters its reasoning phase, its microVM can be marked idle, paused and later resumed from a snapshot. The host can then use physical resources for other active work. For teams running many concurrent sessions, the ability to reclaim capacity during idle periods could matter as much as raw startup speed.
Security is built into the runtime boundary. Agent Compute Engine runs with Titanium, Google’s offload infrastructure processing unit, and uses a Cloud Hypervisor boundary with a separate guest kernel and root access for each microVM. Each environment receives its own agent identity and cryptographically verified ID. Short-lived credentials and service credentials injected outside the microVM are intended to limit what untrusted code can reach.
Google also connects the compute layer to Agent Gateway for observability, policy enforcement and security evaluation. Each microVM receives its own Hyperdisk, allowing execution workspaces and agent context to persist across sessions. Enterprise connectivity is integrated with Google Cloud’s identity and security model, so agents can work with enterprise data without requiring that data to be moved into a separate environment. Google describes the architecture in a technical deep dive.
Gautham Thambidorai, VP and Google Fellow for Gemini, said the service’s microVMs provide the responsiveness, hardware isolation and workload density needed for interactive sandboxes at global scale. The product’s near-term availability is still a preview, but its design points to a broader change: agent infrastructure has to account for activity that can pause and resume many times within a single task.
Kubernetes teams get an open runtime option
For organizations that want more control over the execution layer, or already operate Kubernetes across environments, Google is also offering Agent Substrate on Google Kubernetes Engine. Agent Substrate is an open-source runtime that applies an agent-focused execution model to Kubernetes, separating agent execution from machine management.
Like Agent Compute Engine, it can snapshot an idle agent’s active memory, return the underlying machine to a pool and restore the agent when work resumes. Google says the runtime can deliver up to 10 times the compute density of traditional compute while retaining kernel isolation and identity-based network security for each agent. In benchmark validation involving 200,000 agents, it sustained 500 activations per second with suspend and resume latency below 500 milliseconds.
Teams can choose microVM or gVisor isolation. Each agent receives a cryptographically secure identity, and egress proxies can enforce network policies and inject credentials beyond the agent’s reach. Agent Substrate also integrates with Filestore agent volumes, which Google says can resume in under a second and provide a shared workspace for multi-agent workflows.
Portability is a central distinction: Agent Substrate runs on Kubernetes clusters in different compute environments and supports agent frameworks and harnesses including Claude Code, OpenClaw and Hermes. Nous Research is using the architecture for Hermes Enterprise, which is designed to deploy into customer infrastructure with per-agent isolation and extensible access controls. Hervé Bizira, the company’s chief business officer, said the platform layer addresses those requirements while preserving compute resources.
Governance follows agents across runtimes
As enterprises move beyond isolated pilots, their agents may span different models, tools, data sources and execution environments. That creates a control problem: access should remain limited, asynchronous actions need to be traceable, and traffic between agents and services needs consistent policy enforcement. Google’s governance framework brings several existing Gemini platform components together around those needs.
- Agent Gateway mediates client-to-agent and agent-to-service traffic. Its policy controls cover sensitive data, agent attacks, semantic governance and customer-defined rules.
- Agent Registry provides a central record of approved agents, tools, MCP servers, skills, models and endpoints, with runtime access governed by applied policies.
- Agent Identity assigns agents unique IDs based on the open SPIFFE standard for authentication, access control and auditing.
- Agent observability emits OpenTelemetry-compatible metrics, logs and distributed traces using agentic semantic conventions, covering model calls, tool use and agent handoffs through Agent Gateway.
The expanded coverage includes Google Cloud environments such as Agent Runtime, Cloud Run, Agent Sandboxes in Gemini platform, GKE and Agent Compute Engine, which is entering preview. Google says lightweight Envoy-based enforcement points will extend policy controls to multicloud and hybrid settings in a forthcoming preview. Governance for local coding agents, including Claude Code and Antigravity CLI on managed devices, is also planned, using an operating system sandbox and local governance services.
Gateway controls add cost and routing options
Google plans to preview additional Agent Gateway capabilities this quarter. Token management will let teams track consumption across agents and models, attribute costs and set budget guardrails. Model routing policies will direct requests to selected models with cost and performance in mind. Service Extensions bindings will support custom functions and integrations with Google and partner services directly in the request path.
The announced security integrations include delegated user authorization through Ping Identity, SailPoint and Saviynt, alongside OAuth through Agent Identity’s authentication manager. Google lists its integration partners online. Express Link for Cloud Run is intended to simplify agent-to-tool connections: agents declare the services they need, and the platform establishes an authenticated connection with identity token injection.
L’Oréal is applying Google Cloud’s governance tools to its Beauty Tech Agentic Platform. Alexandre Frizzo, AI lead engineer at L’Oréal, said the company wants to expand agent use across teams while maintaining control over agent access, with identity, policy and observability unified across runtimes such as Cloud Run.
The product set gives platform teams a choice between a managed, agent-specific compute service and a customizable Kubernetes runtime, with governance intended to span both and reach other environments. Agent Compute Engine is expected in preview in the coming months; several cross-cloud, endpoint and Gateway features are also described as upcoming previews. For organizations preparing larger agent fleets, the practical question is now how to match execution density and persistence with identity, policy and telemetry that follow each agent from one runtime to another.
Google provides more information about Agent Compute Engine and an interest form. Teams can explore Agent Substrate and its open-source repository, review agent governance and register interest through Google’s form. Google is also presenting demos at AGNTCon + MCPCon.