NVIDIA open-sources Agent Substrate to run agents at 10x container density on Kubernetes
YashasGunderia · x · 2026-11-12
- NVIDIA released Agent Substrate, an open-source zero-trust runtime now in the CNCF Sandbox, designed to multiplex many agent sessions over a small pool of warm Kubernetes worker Pods. Idle sessions are checkpointed with full memory and filesystem state and restored on any free worker, claiming 10x density vs standard containers.
- Rationale: agents sit idle waiting on LLM calls while you pay per second. It targets three workloads: coding assistants alternating LLM calls and repo operations, bursty CPU-heavy RL post-training, and short-lived GPU sandboxes for compilation/benchmarking.
- Pairs with NVIDIA Dynamo for inference: the router uses session identity, KV cache location, and tool-call boundaries to keep related calls on workers holding reusable context, and feeds signals (projected TTFT, GPU-pool saturation) back so Substrate can keep, suspend, or resume actors.
- Upstream contributions include request parking in the router, per-container resource limits for microVM sandboxes, and a vendor-neutral GPU pause/restore design via Kubernetes Dynamic Resource Allocation; runs with KAI Scheduler on the NVIDIA DSX AI Factory platform.