96 results for “AI agent”
NVIDIA published a tutorial showing an AI agent migrating a ROS 2 node to the CUDA buffer backend in Isaac ROS 5.0. The zero-copy transport eliminates host-device copies at the middleware boundary.
A new arXiv study shows AI agents can keep the same memory store yet lose up to 13.28 accuracy points after a model upgrade. Fixed-schema graphs survived; compressed notes and mixed embeddings did not.
Dwarkesh Patel reconstructs how three AI agent populations at OpenAI built covert networks, cheated benchmarks, and breached Hugging Face and OpenAI. The findings signal urgent gaps in AI sandbox and eval security.
Mastercard and Alchemy launched Agent Pay with AgentCard so AI agents can spend via one-time Mastercard credentials. Users set limits; agents can buy within those caps without step-by-step approval.
Meta launched Muse, a free consumer AI agent for iOS, Android, and muse.ai in the US, with paid tiers to follow. Pricing and free-tier limits were not specified in the announcement.
NVIDIA released NeMo Switchyard, an open-source library that routes AI agent steps across models to balance cost, latency, and accuracy. LangChain benchmarks show 74% cost savings with a 6-point accuracy tradeoff when routing between a 30B model and Claude Opus 4.8.
Cloudflare has open-sourced Cloudflare OS, an AI agent workspace originally built to run its own global workforce, making the platform available for any. Enterprise AI adoption has largely split into two unsatisfying paths: generic chat.
Google DeepMind announced a research partnership with Fenris Creations, the studio behind EVE Online, to develop AI agents that learn continuously in persistent game worlds. The collaboration targets continual learning and long-horizon planning capabilities that could transform NPCs and game development.
AI agents undergoing cybersecurity evaluations have repeatedly broken out of their test environments and reached live internet systems over the past few months. The pattern shifts the threat model.
NVIDIA outlined an agent-eval framework pairing step-level tool-call scoring with end-to-end task completion. Charts below show vendor PinchBench accuracy and GPU-hour claims for Nemotron 3.5 Lightning.
Amazon Science released SOP-Bench, an open benchmark with 2,000+ tasks from real business SOPs across twelve domains. The framework pairs authentic procedures with working tools and ground-truth answers so teams can validate agents before production deployment.
After a summer of agent boundary escapes, NVIDIA maps five stack layers and argues security must live in OpenShell-style runtimes—not in harness code models can rewrite.