Editorial illustration for Google Cloud Launches Agent Compute Engine: A MicroVM per AI Agent
Google Cloud Launches Agent Compute Engine
Google Cloud used this week's AGNTCon + MCPCon event to introduce Agent Compute Engine, a compute service built around a simple premise: every AI agent gets its own hardware-isolated microVM. No shared credentials, no shared kernel, no peeking at another agent's session. The service isn't a concept pitch. Google says it already runs Gemini as an agent for work tasks, along with Antigravity and internal evaluation and reinforcement-learning pipelines.
The design splits an agent into two pieces. A "harness" acts as the brain, calling the model and issuing commands. A "sandbox" acts as the hands, executing code, shell commands and browser actions inside the microVM, with the harness kept entirely outside that boundary. Each session runs on Google-managed hosts using Titan silicon and the Titanium IPU, with its own guest kernel, root access and Hyperdisk storage.
An open-source sibling called Agent Substrate extends the same architecture to any Kubernetes cluster. Google has already published performance figures for Substrate. For Agent Compute Engine itself, the company has stayed quiet on boot times, density and pricing, with customer preview slated for "the coming months."
Google Cloud has announced Agent Compute Engine, a compute service built specifically for AI agents. Google Agent Compute Engine gives each agent its own hardware-isolated microVM. The sandbox holds zero credentials.
Why this matters
The isolation pitch is the real story here. Running agents in credential-free microVMs rather than shared containers addresses a problem most teams have been papering over: what happens when an autonomous agent with broad permissions gets prompt-injected or just goes off-script. Hardware-level sandboxing, paired with zero stored credentials, is a sane default for anyone handing agents real tasks instead of demo scripts.
The 10x density claim and 500 activations per second under 500ms matter too, because the usual argument against VM-grade isolation has always been cost and latency. If Google can actually hold those numbers at scale, that argument weakens considerably.
The open-source Agent Substrate release on GKE is the part worth watching closely. A Kubernetes-native layer with Cloud Hypervisor and gVisor options means this isn't locked to Google's own agent stack, so builders can test the architecture without committing to Agent Compute Engine itself. We'd treat the preview timeline and those performance figures as claims to verify once third parties get hands-on access, not conclusions.
Related Reading
Browse all Google news →
Business & Startups Nordic pilot adds Gemini for Education, NotebookLM to boost AI literacy
Business & Startups Kling launches Video O1, all-in-one model with MVL bridge using transformer
Business & Startups DeepSeek Seeks More Capital Weeks After USD 7B Funding Round
Related Topic NVIDIA and Google Cloud let developers scale AI from prototype to production
Market Trends
Related Topic Google tests visual 'magazine-style' UI for Gemini 3 Pro users
LLMs & Generative AI
From Archives Kazakhstan Expands Compute for AI Future as Huang Calls Physical AI Fantastic
From Archives Former OpenAI Exec Raises USD 215M to Build AI Drug Labs Guided by Chemists
From Archives OpenAI brings models and Codex to Google Cloud, Brockman says
From Archives US Apple CarPlay Users Can Now Use Gemini to Chat With Google Maps
LLMs & Generative AI
Common Questions Answered
What is the key security feature of Google Cloud's Agent Compute Engine?
Each AI agent receives its own hardware-isolated microVM with zero stored credentials, eliminating shared kernels and preventing agents from accessing other agents' sessions. This hardware-level sandboxing approach addresses the vulnerability of autonomous agents being prompt-injected or deviating from their intended behavior when given broad permissions.
How does Agent Compute Engine's microVM architecture differ from traditional shared container deployments?
Agent Compute Engine allocates a dedicated hardware-isolated microVM to each agent rather than running multiple agents in shared containers with shared credentials and kernels. This isolation ensures complete separation between agents, preventing one agent's compromise from affecting others and eliminating credential-sharing risks inherent in traditional container-based approaches.
What real-world applications is Google already using Agent Compute Engine for?
Google Cloud is currently running Gemini as an agent for work tasks, along with Antigravity and internal evaluation and reinforcement-learning pipelines on Agent Compute Engine. These production deployments demonstrate that the service is not merely a concept but an actively deployed infrastructure for managing autonomous AI agents.
Why is the credential-free sandbox design important for autonomous agents handling real tasks?
A credential-free sandbox prevents compromised or misbehaving agents from accessing sensitive authentication information that could be exploited for unauthorized actions. This design is particularly critical for agents handling real business tasks rather than demos, as it provides a sane security default that limits the blast radius of agent failures or malicious prompt injections.
Further Reading
- Inside Agent Compute Engine: Building the compute foundation for the agentic era - Google Cloud Blog
- Introducing Agent Compute Engine and expanding agent governance - Google Cloud Blog
- Google Cloud Adds Isolated MicroVMs for AI Agents, Preview Coming - TokenPost
- Google Unveils Agent Compute Engine, a microVM for Every AI Agent - Pasquale Pillitteri
- Google Cloud Launches Agent Compute Engine: A MicroVM per AI Agent - MarkTechPost
