Agent Compute Engine

Google's new service runs AI agents in isolated microVMs with secure checkpointing

Google's Agent Compute Engine runs AI agents in isolated microVMs with checkpointing, letting you suspend and resume execution environments without exposing credentials.

3 min readInfoQ
Google's new service runs AI agents in isolated microVMs with secure checkpointing

The most important thing Google announced with Agent Compute Engine isn't the ability to run AI agents in isolated microVMs, it's that they finally took secrets management seriously. For too long, the industry has treated credential storage as an afterthought in agentic workflows, and the results have been predictable. A recent investigation into AI coding tool Claude Code exposed 2,282 secrets in public repositories showed exactly what happens when agents are given broad access without architectural guardrails. Google's approach, keeping API keys, OAuth tokens, and credentials entirely outside the agent sandbox, routed through an external Agent Gateway, isn't just clever engineering. It's the minimum viable architecture for any agent system that touches production data.

The design matters because it addresses a tension that has been building across the ecosystem. Agents need to act on our behalf, but every action they take with stored credentials is a potential blast radius. By introducing secure checkpointing that can suspend and resume execution environments, Google is acknowledging that long-running agents are inevitable. After all, when 1,000 AI agents ran for 8 days, one insight stood clear: reliability and state management are the real bottlenecks, not agent intelligence. Checkpointing solves the state problem, but only if the secrets that agents rely on survive those restarts securely. Google's gateway model ensures that credentials are injected at runtime rather than persisted in the agent's memory, which means a checkpointed agent can be restored without carrying forward stale or compromised tokens.

For teams building production agents today, this is the architectural pattern worth studying. The microVM isolation prevents one agent from leaking into another, and the external gateway means that even if an agent is compromised, the attacker doesn't walk away with your API keys. That's a meaningful improvement over the current default, which is essentially trusting agents to manage their own secrets, a pattern that has already proven fragile. The real test will be how this service handles the messy reality of multi-step workflows that require different credentials at different stages, and whether the checkpointing mechanism can preserve enough context to make resumption feel seamless rather than like a cold restart.

What remains an open question is how this will scale when agents need to collaborate with each other or with external services that have their own authentication flows. The gateway approach works well for a single agent talking to a fixed set of APIs, but agent-to-agent delegation introduces new trust boundaries that the current design doesn't fully address. That's the detail to watch: how Google handles cross-agent credential handoff without reintroducing the very vulnerabilities the isolation model was built to eliminate.

From InfoQ

Google recently announced Agent Compute Engine, a managed service that runs AI agents in isolated microVMs, with checkpointing capabilities to suspend and resume execution environments. The design keeps credentials, API keys, and OAuth refresh tokens outside the agent sandbox, using an external Agent Gateway to authenticate requests and inject the required secrets.

Read the original at InfoQ