Google Launches Agent Compute Engine for AI Agents
The managed service runs AI agents in isolated microVMs, keeping credentials outside the agent sandbox.
Updated on Nov. 22, 2026 in Artificial Intelligence

Google announced Agent Compute Engine, a managed service that runs AI agents in isolated microVMs with checkpointing to suspend and resume execution environments. The service is not yet available and is limited to Google Cloud.
Why it matters
Google designed the service to keep sandboxes free of exploitable credentials, even when agents make mistakes. The announcement arrives as cloud providers compete to offer managed environments for running autonomous AI agents.
Agent Compute Engine runs multiple agents on a single operating system template using copy-on-write memory sharing, and idle microVMs are paused via fast checkpoint and restore. Google has not published benchmark figures for these mechanisms, and the preview timeline remains a targeted goal rather than a confirmed date.
The players
Google is the Alphabet subsidiary whose cloud division operates Google Cloud and is developing the Agent Compute Engine and Agent Substrate runtime.
The details
An external Agent Gateway authenticates requests and injects required secrets, while the orchestration harness runs outside the execution sandbox and dispatches explicit commands to untrusted guest environments. Google also offers Agent Substrate, an open-source Kubernetes-native runtime, alongside a governance stack that includes Agent Registry, SPIFFE-based Agent Identity, and OpenTelemetry-compatible observability.
Timeline
Google targets a preview of Agent Compute Engine in the coming months.
The Tech Race
Agent Compute Engine enters a market where AWS Bedrock AgentCore and Azure Foundry Agent Service already provide managed isolated execution for AI agents. Google's entry positions it against those rival cloud platforms in the race to host autonomous agent workloads.
Developers building AI agents on Google Cloud could gain a way to run them without storing credentials inside the sandbox, reducing the risk of exposed keys. Availability is limited to Google Cloud and the service is not yet available.
The takeaway
Keeping credentials outside an agent's execution environment is a design choice that limits the damage a faulty agent can cause. Teams evaluating managed agent platforms should compare how each provider isolates secrets before committing workloads.
What happens next
Google has targeted a preview of Agent Compute Engine in the coming months, though no specific date has been set.
Further reading
Explore more on Artificial Intelligence developments shaping cloud computing.
Source note: This article includes information reported by InfoQ.






