NVIDIA AI Updates: November 5, 2026
Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled
1. CUDA 13.5 Makes Vera Rubin Support Generally Available and Adds MPS v3 Scheduling Controls and Multi-GPU Communication in CUDA Tile
NVIDIA released CUDA Toolkit 13.5, which moves Vera Rubin support from early developer access to general availability and extends Windows on Arm support (GA since 13.4) to DGX Station for Windows. MPS v3 adds weighted timeslicing per MPS server, SM partition load balancing (borrowing with --max-chunks, lending idle SMs with --shareable=true), cgroup v2 routing of clients into MPS namespaces, per-namespace SM partition assignments that replace per-client CUDA_MPS_SM_PARTITION settings, and a polling fallback for memory accounting, while CUDA Graphs can now capture batched 1D copies made with cudaMemcpyBatchAsync. CUDA Tile kernels gain multi-GPU communication over NVLink (unicast pointers, plus multicast pointers and multimem reductions on NVSwitch fabrics with NVLink SHARP), and a public preview lets cuTile Python and CUDA Tile C++ kernels call SIMT library device functions such as NVSHMEM, CCCL Coop, Warp and MathDx mid-kernel instead of only across kernel boundaries. CCCL 3.5 adds public CUB tuning APIs, run-to-run reproducible floating-point scans and cub::DeviceBatchedTopK for per-segment top-K selection (already integrated into FlashInfer); cuBLAS improves out-of-the-box NVFP4 and MXFP8 matmuls on Vera Rubin, including MoE training shapes; and the toolkit adds a public preview of the Nsight Perf SDK for AI for always-on profiling of long-running training and inference jobs. Two compatibility changes: Rubin and later GPUs do not support Tesla Compute Cluster (TCC) mode on Windows, only WDDM and MCDM, and cuobjdump symbol-table output now uses four named columns, so scripts that parse it need updating. Source
2. NVIDIA Releases the Multi-Agent Framework Behind Its Internal DGX Cluster Operations as Experimental Open Source
NVIDIA published the AI Factory Operations Agent, a multi-agent system it built for the SREs running its internal DGX H100, DGX B200 and DGX GB300 production clusters, as a developer example plus an Apache 2.0 GitHub repository that NVIDIA labels experimental, with APIs and deployment requirements subject to change. Built on the NemoClaw blueprint with locally served Nemotron models, a central agent routes chat queries to domain agents for job recovery (training failures and goodput drops), hardware faults (PCIe, ECC, NVMe and thermal), observability (Prometheus and Grafana Loki), documentation research, read-only Kubernetes state and Base Command Manager diagnostics, each reaching infrastructure through MCP servers that expose a fixed set of approved tools. The system deploys with Helm as a Kubernetes control-plane application and runs shell commands in a per-session OpenShell sandbox, while Kubernetes inspection goes through a server-side run_kubectl tool with a pinned binary and read-only credentials that the sandbox never receives, so RBAC lets it read pods and configmaps but not secrets, and not create or update pods. NVIDIA lists five lessons from production use (non-overlapping agent boundaries, read-only investigation before automation, cross-layer correlation, precise MCP tool definitions and visible reasoning) and recommends running the agents on a dedicated GPU control plane isolated from the production cluster, with an RTX PRO Workstation or DGX Spark for smaller sites. The orchestration framework is open source, while the NVIDIA-developed cluster operations agents are available to customers on request through their account managers. Source
3. NeMo Compass and NeMo Eval Author Turn Agent Traces Into a Ranked Issue List and Custom Evaluations
NVIDIA published a tutorial for its NemoClaw blueprint for agent optimization that introduces two Apache 2.0 libraries: NeMo Compass, which reduces a corpus of agent traces and evaluation evidence to a prioritized backlog of issues linked to trace IDs, and NeMo Eval Author, a set of coding agent skills (for agents such as Codex or Claude Code) that generates Harbor or NeMo Gym evaluations and verifiers from those findings. In the worked example, Compass compared 42 traces from a Hermes Agent enterprise assistant against an ETHOS.md file of expected behavior and found the agent calling send_message without user confirmation; after a one-paragraph change to the agent’s SOUL.md instructions, pass rates rose from 3/6 to 6/6 on development tasks and from 4/6 to 6/6 on held-out tasks kept hidden from the coding agent, with tool calls falling from 10 to 6 and from 28 to 15. In a second example with Dassault Systèmes’ LEO engineering companion on a car door design workflow, an overnight loop of about 12 hours without human intervention cut malformed QueryModel tool calls from 142 to zero and reported a 28% relative improvement in the overall evaluation score with 35% fewer tool calls per run across six scenarios. The setup requires a Linux machine running Docker natively (Ubuntu 24.04 with 8 vCPUs and 32 GB of RAM recommended), OpenShell 0.1.2, Harbor 0.22.0 and an NVIDIA API key. Source
4. NVIDIA’s Factory Operations Blueprint (FOX) Opens Early Access for Manager Agents That Coordinate Inspection, Rework and Root Cause Analysis
NVIDIA detailed the Factory Operations Blueprint (FOX), a reference architecture, available by early access request, for adding a long-running factory manager agent built with the NVIDIA Agent Toolkit on top of existing machine databases, MES and ERP systems. The manager can run on a frontier model such as GPT-6 Astra or an open model such as Nemotron, uses NemoClaw as its default harness, and executes inside OpenShell, where factory policies define permitted file, command and network actions and which steps need operator review; it delegates to specialist agents that call Cosmos for visual inspection and Kumo Anomaly for time-series anomaly detection, with Elasticsearch holding production evidence, versioned manuals and prior investigations. The reference example follows an automated optical inspection alert on a surface-mount PCB line: a Cosmos-based agent rechecks the board against handbook examples using in-context learning, a fleet management agent moves confirmed defects to rework in an Isaac Sim and ROS 2 digital twin, and a root cause agent pulls the board’s reflow oven readings and returns a PDF report with suspect measurements and a recommended corrective action. FOX can run fully on premises on a DGX Station, and NVIDIA recommends operator approval before any physical handling in initial deployments. Separately, Siemens uses FOX-built, Nemotron-powered agents at its Erlangen electronics factory to run rush-order planning simulations through Siemens Intelligence Center X, which NVIDIA says can enable cost savings of up to 20%. Source