An agent is only as good as the harness around the model. At NVIDIA, Matt's team built NOOA (NVIDIA Labs Object-Oriented Agents), an open-source research preview that treats the agent as a Python object: typed input/output, pass by reference on live objects, code as action, programmable loop engineering, explicit object state, and model-callable harness APIs.
Matt walks through these six capabilities, shows how NOOA runs inside NVIDIA OpenShell, and what they buy you in numbers: 82.2% on SWE-bench Verified, 50.2% mean RHAE on ARC-AGI-3 and 86.8% on CyberGym L1 - the top-scoring open-source agent there, at roughly half the token cost of comparable harnesses.
Put your tools, products, and engineering work in front of builders working with agentic AI.