NVIDIA pushes OpenShell and its open-source tools to govern AI agents
At KubeCon, NVIDIA donated OpenShell to the CNCF and released several open source components for its DSX AI Factory Platform. Goal: better secure, orchestrate, and restore AI agents on Kubernetes.
On November 10, 2026, at KubeCon + CloudNativeCon, NVIDIA announced the donation of OpenShell to the Cloud Native Computing Foundation, while releasing three open source components for its DSX AI Factory Platform. The challenge goes beyond a simple community gesture: it is about making AI agents that handle tools, consume compute, and sometimes need to restart without losing everything, safer. In this context, the company links governance, Kubernetes orchestration, and disaster recovery in a single technical package.
Des agents mieux encadrés, enfin sans magie noire
OpenShell, a security runtime for AI agents, applies deterministic controls when autonomous agents access tools, inference services, or execute workloads. The engine runs in kernel-level isolated sandboxes, with declarative YAML policies covering access to the file system, network, processes, and provider credentials. Credentials are injected only into requests destined for approved endpoints. NVIDIA also adds formal verification to test whether a policy change would open risky access before approval. The idea is simple, even if the tooling is not: prevent an overly curious agent from doing what it shouldn’t.
The project is distributed under the Apache 2.0 license. NVIDIA says hundreds of organizations have already used it and tens of thousands of contributions have arrived in less than a year. The transfer to the CNCF is still ongoing: a project notice indicates that development is migrating to the ai-openshell organization, while code, tickets, and pull requests remain with NVIDIA/OpenShell during the transition.
Agent Substrate veut éviter le coût des agents qui attendent
Agent Substrate tackles a very concrete problem: agents often remain idle waiting for an instruction, a tool response, or a model result, but still consume memory and CPU. According to NVIDIA, giving each agent its own Kubernetes Pod blocks resources and inflates the number of persistent objects to track. The project therefore aims to decouple stateful agent sessions and their identity from active compute, then restore these sessions in memory when needed, with a restoration time announced under 500 milliseconds.
The mechanism relies on a small number of “hot” workers hosting multiple sessions. When an agent becomes inactive, its state — including memory and file system — is checkpointed, the worker is freed, then the agent is restored on an available resource without going through the usual Pod creation. Workers remain ordinary Kubernetes Pods, preserving usual scheduling, quota, and placement rules. Target use cases include code assistants, reinforcement learning environments, and short-lived GPU sandboxes.
Open source, mais pas encore figé
NVIDIA also plans interoperability with Dynamo for inference and KAI Scheduler for cluster resource sharing, two components linked to its DSX AI Factory Platform. Dynamo can provide an estimate of time before the first token and GPU pool saturation, so Agent Substrate can decide to keep an agent in memory, suspend it, or restart it. APIs are still evolving, GPU support is being rebuilt around Kubernetes Dynamic Resource Allocation, and part of the integration remains experimental.
In other words, the architecture already exists, but it does not yet have the stability of a well-established product. Open source does not mean ready to plug in everywhere the next morning, even if the temptation is strong.
Snapshot promet de réduire le temps de reprise à quelques secondes
NVIDIA Snapshot targets another friction point: cold start of inferences. The tool checkpoints a fully initialized workload, including CPU and GPU memory, then restores it without reloading weights, compiling, or capturing CUDA graphs. Restoration requires the same GPU model and driver, which limits the scope but greatly speeds up recovery.
On B200 GPUs with VAST NFS storage, NVIDIA reports restorations from 3.6 to 42.7 seconds depending on vLLM, SGLang, and TensorRT LLM configurations, representing gains of 8x to 36x compared to cold start. Tests covered a 600 million parameter model up to eight-GPU configurations with checkpoints approaching 1 TB. Creating a checkpoint took between 15 seconds and 5.5 minutes, and the combined checkpoint + restore time remained below cold start in all tested cases.
These measurements do not cover the entire machine startup: infrastructure provisioning and container image loading remain outside the timed restore path. Another important detail: the key-value cache is released before each checkpoint rather than kept in the artifact, meaning results concern resuming an already initialized workload on compatible infrastructure.
Les gains dépendent aussi du stockage, pas seulement du GPU
Crusoe also tested its own GLM-5.3 use case on NVIDIA HGX B300 with a 594 GB GPU state checkpoint. The company measured GPU state restoration in 15.0 seconds from local NVMe, versus 33.3 seconds from NFS. With the artifact on NVMe, Pod-to-ready time was 31 seconds, versus about eight minutes for a cold start. This comparison shows that storage heavily impacts final performance. Crusoe also contributed performance and diagnostic improvements upstream.
Taken together, these projects shift more execution control, security, and recovery towards open repositories. But they do not all progress at the same pace: OpenShell is in transition, Agent Substrate continues to evolve, while NVKE and NVKR remain in public preview. The promise is clear; long-term maintenance will depend mainly on tooling, APIs, and compatible hardware.
Key points
- OpenShell joined the CNCF Sandbox on November 10, 2026.
- Agent Substrate targets restorations under 500 milliseconds.
- Snapshot reports 3.6 to 42.7 seconds on B200.
- Crusoe measures 31 seconds with local NVMe.
- Mirantis and SK Telecom are among cited partners.
- NVKE and NVKR remain in public preview.
By the numbers
- November 10, 2026 — announcement made at KubeCon + CloudNativeCon.
- 3.6 to 42.7 seconds — Snapshot restorations on B200 with VAST NFS.
- 8x to 36x — gains announced versus cold start.
- 594 GB — size of checkpoint tested by Crusoe on HGX B300.
- 31 seconds — Pod-to-ready time with local NVMe at Crusoe.