Nvidia has released DeepStream 9.1, an update to its software toolkit for building video analytics systems, with a focus on making these tools accessible through natural language and AI coding assistants.
According to MarkTechPost, the headline feature is a set of 13 "agentic skills" that let AI coding agents — including Anthropic's Claude Code and OpenAI's Codex — assemble multi-camera video analytics pipelines from plain-language prompts. In practice, that means a developer could describe what they want a network of cameras to do, and a coding agent could wire up the underlying system rather than a human hand-coding every step.
The release also introduces Multi-View 3D Tracking, or MV3DT. As described by MarkTechPost, this feature fuses detections from individual cameras into a single shared 3D representation of a space, assigning each tracked object one globally consistent ID across every camera. So a person or vehicle moving through a building covered by many cameras would keep the same identity as they pass from one camera's view to the next, instead of being counted as a new object each time.
DeepStream is Nvidia's framework for so-called "vision AI" — software that analyzes live video feeds for uses such as security, retail analytics, traffic monitoring and industrial inspection.
Why it matters: By letting AI agents build camera-analytics pipelines from natural language and by stitching many camera views into one coherent 3D picture, DeepStream 9.1 signals how quickly the agentic-AI trend is spreading beyond chatbots into the specialized software that powers real-world surveillance and monitoring systems.