NVIDIA AI Updates: August 21, 2026

1. NVIDIA SkillEvaluator Measures AI Agent Skill Performance

NVIDIA. NVIDIA introduced SkillEvaluator, a tool for measuring the performance of AI agent skills. The article outlines how developers can benchmark and assess agent capabilities across tasks rather than relying on ad hoc testing. It targets a gap in reproducible evaluation as teams ship more agentic systems. Source

2. Post-Training Cosmos 3 Edge for On-Device Robot Control

NVIDIA. NVIDIA published a walkthrough for post-training the Cosmos 3 Edge world foundation model for on-device robot control. It covers adapting the model for low-latency robotics inference at the edge, where connectivity and compute are constrained. The guide is aimed at teams deploying manipulation and control policies directly on robots. Source

3. Generative Recommenders for Large-Scale RecSys

NVIDIA. An NVIDIA developer post details how generative recommender architectures are being applied to large-scale recommendation systems. It covers techniques for handling ranking and retrieval workloads at production scale using NVIDIA’s tooling. The framing reflects a shift from traditional pipelines toward generative approaches in recommendation. Source

4. Building Federated Multimodal AI Workflows With NVIDIA FLARE

NVIDIA. NVIDIA describes how to construct federated multimodal AI workflows using the NVIDIA FLARE framework. The article addresses training across distributed data sources while preserving data privacy, a common requirement in healthcare and finance. It provides a reference architecture for combining modalities without centralizing sensitive data. Source

5. Developing Holoscan Applications With CLI, Skills, and AI Coding Agents

NVIDIA. NVIDIA explains how to build Holoscan applications using the Holoscan CLI alongside skills and AI coding agents. The post focuses on streamlining development workflows for real-time medical and edge AI applications. It connects Holoscan’s sensor-processing pipeline tooling to agent-assisted development. Source