Google has announced Agent Compute Engine, a Google Cloud service designed to give every AI agent its own isolated virtual computer. The platform is built to let agents run code, operate browsers, use software tools and maintain working environments across multi-step tasks without sharing a conventional execution environment with other agents.
Announced in October 2026, Agent Compute Engine addresses a growing infrastructure challenge as AI systems move beyond generating text to taking actions independently. Google says the platform already supports internal workloads, including Gemini, and is expected to enter preview on Google Cloud in the coming months. The service combines hardware-isolated micro virtual machines, persistent workspaces and automated resource management to help developers run agents more securely and efficiently.
Key Takeaways
- New service: Agent Compute Engine gives AI agents dedicated, hardware-isolated microVM environments.
- Security: Agents can execute code inside isolated environments, with controlled network access and safeguards designed to keep long-lived credentials outside the sandbox.
- Lower costs: Idle environments can be suspended while preserving their working state, reducing the need to reserve computing resources continuously.
- Developer use cases: The service targets coding agents, enterprise automation, browser-based tasks and large-scale AI evaluations.
- Availability: Google plans to make the platform available in Google Cloud preview in the coming months; it is not yet a generally available service.
What Is Google Agent Compute Engine?
Agent Compute Engine is a cloud infrastructure service built specifically for AI agents. Instead of treating an agent as a short-lived process that runs inside a shared application environment, Google gives it a computer-like workspace where it can execute commands, access files, use programming tools and continue work across multiple interactions.
This matters because AI agents need more than a language model. A coding agent, for example, may need a Linux environment, a file system, compilers, a terminal, a browser and network access. An enterprise agent might need to inspect documents, query databases or interact with business applications.
Running these activities safely at scale can be difficult with conventional virtual machines or containers. Traditional virtual machines provide strong isolation but can be comparatively slow and resource-intensive to start. Containers are lightweight and fast, but share the host operating system’s kernel, which may not provide the isolation required for some untrusted workloads.
Agent Compute Engine is Google’s attempt to combine the isolation of virtual machines with the speed and efficiency needed for large numbers of AI agents.
How the Isolated Virtual Machine System Works
The platform uses micro virtual machines, or microVMs, to create separate execution environments for agents. These environments can run code and tools without giving every agent unrestricted access to the underlying host system.
Google describes several key components of the design:
- Dedicated environments: Each agent session runs inside its own isolated microVM.
- Persistent state: Files and working state can survive pauses and resume across later interactions.
- Controlled networking: Agent traffic passes through managed boundaries rather than relying on unrestricted direct network access.
- Credential protection: The sandbox is designed not to store long-lived credentials, such as persistent cloud keys or OAuth refresh tokens.
- Security monitoring: Google describes host integrity checks, encryption, vulnerability scanning and runtime protections as part of its security approach.
The separation between the agent’s orchestration system and its execution environment is especially important. The orchestration layer decides which task to run, while the sandbox executes the requested commands. This helps limit the damage if an agent encounters malicious instructions or runs untrusted code.
Isolation does not make an agent automatically safe, however. Developers still need to configure appropriate permissions, network restrictions and policies for the systems their agents can access.
Why Google Wants AI Agents to Have Their Own Computers
AI agents can behave differently from traditional applications. An agent may run a command, wait for a model to generate its next step, execute another tool and then pause until a user responds. During these waiting periods, keeping a complete virtual machine actively allocated can waste resources.
At the same time, repeatedly shutting down an environment can destroy the agent’s working state and slow down the next interaction.
Agent Compute Engine is designed to address both problems by suspending idle environments and restoring them when work resumes. Google says its infrastructure can preserve an agent’s state while reducing active resource consumption during idle periods.
The platform also uses snapshots and memory deduplication. When multiple agents use the same base operating system, libraries or tools, Google can share unchanged memory pages rather than keeping identical copies for every environment.
Together, these mechanisms are intended to reduce infrastructure costs and make it practical to operate very large numbers of persistent agents.
What Developers Can Use Agent Compute Engine For
Google identifies several categories of workloads that could benefit from the service.
| Use case | How the infrastructure helps |
|---|---|
| AI coding agents | Provides persistent Linux workspaces, compilers, terminals and software-testing tools |
| Enterprise automation | Supports agents that interact with documents, databases and business systems |
| Browser-based agents | Creates isolated environments for agents that navigate websites and perform tasks |
| AI model evaluation | Enables many parallel test environments to evaluate model behavior |
| Reinforcement learning | Supports repeated simulations and experiments in separate environments |
For software developers, persistent workspaces could be particularly useful. An agent might modify a codebase, run tests, pause overnight and resume the next day without rebuilding its environment from scratch.
For enterprises, isolated execution environments could help teams experiment with agents that process internal information or execute commands. Those deployments would still require appropriate identity management and permission controls.
Security Is a Central Part of the Design
The announcement arrives as companies explore agents capable of making decisions and acting on users’ behalf. These capabilities create security challenges that go beyond the risks associated with ordinary chatbots.
For example, an agent reading an email or document could encounter malicious instructions designed to make it expose private information or send data to an external destination. A model may not reliably recognize every such attempt.
Google’s approach is to make the execution environment safer even when the model makes a mistake. Its design includes restricted network access, short-lived session identities and a separation between the system that authorizes actions and the sandbox that executes them.
This is a useful security principle: developers should not depend entirely on an AI model to follow instructions correctly. They should also limit what the model’s tools can access and enforce important rules outside the model itself.
However, a dedicated microVM is only one layer of protection. An agent can still misuse permissions intentionally or accidentally granted to it. Businesses must configure the right access scope, monitor activity and require human confirmation for sensitive operations where appropriate.
How Agent Compute Engine Could Change AI Infrastructure Costs
Running AI agents at scale creates a different cost profile from serving a conventional chatbot. A single request may trigger several model calls, browser actions, code executions and background tasks. If every agent requires a fully active virtual machine throughout the process, costs can rise quickly.
Agent Compute Engine is designed to reduce this overhead by releasing active resources when agents are idle and sharing identical memory pages across environments.
Google says its internal systems are built to support workloads ranging from interactive agents to large-scale evaluations. Independent reporting has also highlighted the platform’s focus on rapidly suspending and resuming agents, although performance and savings for customers will depend on their workloads and configurations.
The commercial impact will become clearer once preview access allows developers to test the service and compare its costs with existing infrastructure. Google has not yet established a generally available release date in the announcement.
The Bigger Picture
Agent Compute Engine reflects a broader shift in AI infrastructure. As AI systems gain the ability to execute code, browse the web and interact with business applications, the cloud must provide environments that are not only powerful but also isolated, persistent and economical to operate.
Google is positioning agent-specific infrastructure as a foundation for the next stage of AI development. The competitive opportunity is not limited to offering more computing power: it also involves making autonomous software easier to deploy, control and scale. If the approach works as intended, developers may spend less time building custom sandboxes and more time designing useful agent workflows.
FAQs
1. What is Google Agent Compute Engine?
Agent Compute Engine is a Google Cloud service that provides isolated micro virtual machines for AI agents to execute code, use tools and maintain persistent working environments.
2. How is it different from a regular virtual machine?
Traditional virtual machines are general-purpose computing environments. Agent Compute Engine is designed around agent-specific needs, including fast suspension and resumption, persistent state, resource efficiency and controlled execution.
3. Is Agent Compute Engine available now?
Google says the service is expected to enter Google Cloud preview in the coming months. The October 2026 announcement does not establish a general availability date.
4. Which AI agents could benefit from the service?
Potential users include coding agents, enterprise workflow agents, browser-based agents and systems that run large-scale AI evaluations or reinforcement-learning experiments.
5. Does an isolated virtual machine make an AI agent completely secure?
No. Isolation can reduce risks from code execution and help contain an agent’s environment, but security also depends on identity management, permissions, network policies, monitoring and the design of connected tools.
Looking Ahead
The next milestone will be the preview release, when developers can assess how Agent Compute Engine performs in real applications and whether its security and resource-management features deliver practical benefits. Pricing, service limits and availability details will be important factors for teams deciding whether to adopt it.
As AI agents become more capable, infrastructure providers will increasingly compete on secure execution, reliable access controls and the cost of running autonomous workloads. Google’s new service is an attempt to make those capabilities part of the cloud platform itself, rather than leaving every developer to build the underlying systems independently.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
