Google has donated Agent Substrate and gVisor to the Cloud Native Computing Foundation, placing two parts of its AI-agent infrastructure under vendor-neutral governance. It is also contributing OpenRL, an orchestration project for reinforcement-learning workloads.
Agent Substrate addresses the stop-and-start pattern of agents that execute briefly, then wait for a model, external API or human approval. Google estimates such agents may sit idle 90 to 95 percent of the time while retaining files, memory and active sessions. The system stores that state as an “actor” and assigns execution to pre-warmed shared workers only when work resumes.
Google says the approach can restore an agent in under 500 milliseconds without creating a new Kubernetes Pod for every step. It also reduces pressure from thousands of short-lived environments during large evaluation runs. That is a vendor claim, and operators will need to benchmark recovery time and isolation under their own workloads.
gVisor supplies a separate security layer by intercepting system calls in a user-space kernel rather than exposing the host Linux kernel directly. Agent Substrate can also use microVMs for workloads needing hardware virtualization. The combination is designed for generated shell commands and packages that cannot be treated as trusted, while allowing platform teams to choose a stronger boundary for higher-risk work.