NVIDIA Open-Sources NOOA, an Agent Framework Where the Agent Is One Python Class and the LLM Fills In Method Bodies Marked With Ellipses
NVIDIA Labs released Object-Oriented Agents (NOOA) as an open research preview that collapses prompts, tools, callbacks and workflows into a single Python class: state lives in fields, capabilities in methods, prompts in docstrings, and type annotations act as enforced contracts, with LLM-driven loops completing any method body left as an ellipsis at runtime. NVIDIA reports state-of-the-art accuracy and lower token cost versus prior harnesses on SWE-bench Verified, CyberGym L1 and ARC-AGI-3, and released the framework, capability tests, benchmark agents, evaluation methodology and a technical report (arXiv 2607.20709) together. The repo at NVIDIA-NeMo/labs-OO-Agents is at 490 stars with commits through July 29, and the practical appeal is that agents defined this way stay compatible with ordinary Python testing, tracing, refactoring and version control.
↳ Follow the thread