Stacks/Agent platform / infrastructure

NVIDIA Agent Substrate

NVIDIA's Agent Substrate is a production-oriented Kubernetes runtime, donated to the CNCF Sandbox, for the idle-heavy shape of agent workloads. Many sessions share a pool of warm Pods; idle ones are snapshotted (memory and filesystem) and restored later. It pairs with NVIDIA Dynamo for inference routing and KAI Scheduler for cluster placement, and targets coding assistants, RL environments, and GPU sandboxes.

Ekin Karabulut and Omer YahudNov 9, 2026Original writeup ↗
Use case
A security-hardened, zero-trust Kubernetes runtime for running sandboxed agents and tools at high density — coding assistants, RL post-training environments, and short-lived GPU sandboxes — now donated to the CNCF Sandbox
Category
Agent platform / infrastructure
Team size
not stated
Orchestrator / framework
Agent Substrate multiplexes many agent sessions over a small pool of warm Kubernetes worker Pods. A separate control plane maps actors (agents and their environments) onto that pool. NVIDIA Dynamo coordinates inference across GPU workers; Dynamo's router uses session identity, KV-cache location, and tool-call boundaries to keep related model calls on workers that already hold reusable context. KAI Scheduler places Substrate worker Pods alongside GPU trainers and inference servers on the NVIDIA DSX AI Factory Platform.
Memory
Idle sessions are checkpointed, memory and filesystem included, and restored onto whichever worker is free when the next request arrives. Dynamo exposes inference signals (projected time to first token, GPU-pool saturation) so Substrate can keep an actor resident, suspend it during long inference waits, or resume it as the response nears completion.
Evals
not stated
Observability
not stated
Monthly cost
not stated
Models
not stated
Observability tools
not stated
Other tools
gVisor or microVM sandboxesKubernetes Dynamic Resource Allocation (for pausing and restoring GPU state)CNCF Sandbox

Related

More like this