NVIDIA donates Agent Substrate runtime to CNCF
NVIDIA has given Agent Substrate, an open-source Kubernetes runtime for sandboxed agents, to the CNCF Sandbox, targeting the idle-heavy workloads of coding agents and RL post-training that standard Pods handle poorly.
NVIDIA announced that it has donated Agent Substrate to the Cloud Native Computing Foundation (CNCF) Sandbox. NVIDIA calls it a security-hardened, zero-trust runtime built to run millions of agents and tools at 10x the density of standard containers. That figure is NVIDIA's own and has not been independently verified.
The problem it targets is that agents spend most of their time idle, waiting on a user, a tool or a model. On Kubernetes, teams usually choose between keeping a warm Pod for each session, which wastes compute, and creating Pods on demand, which means a cold start of several seconds at every step.
Agent Substrate keeps a pool of ordinary Kubernetes worker Pods and adds its own control plane. That control plane maps a much larger set of "actors" (agents and their environments) onto those workers. When an actor goes idle, its sandbox (gVisor or a microVM) is snapshotted, including memory and filesystem, and the worker is freed. When a new request arrives, the actor is restored onto any free worker without going back through the API server, scheduler or kubelet.
NVIDIA lists three target workloads:
- coding-agent serving
- RL post-training, where it says sandbox fleets can grow to "many hundreds of thousands" and slow environments leave GPUs idle
- short-lived GPU sandboxes for research
Recent contributions include:
- Request parking (#221): when every worker is busy, requests wait instead of failing.
- Per-container CPU and memory limits for microVM sandboxes (#859).
- GPU support through Kubernetes Dynamic Resource Allocation (#752): in progress, with NVIDIA GPUs supported first.
NVIDIA also says the project works with KAI Scheduler and NVIDIA Dynamo in its DSX AI Factory Platform. The project is still early and changing.