NVIDIA's NOOA makes the agent one Python class with ellipsis method bodies.

NVIDIA Labs released Object-Oriented Agents as a research preview that collapses prompts, tools, callbacks, and workflows into a single class: state in fields, capabilities in methods, prompts in docstrings, type annotations as enforced contracts, and LLM-driven loops completing any method body left as ... at runtime. They report SOTA accuracy and lower token cost on SWE-bench Verified, CyberGym L1, and ARC-AGI-3, with the framework, capability tests, benchmark agents, and a technical report (arXiv 2607.20709) all released together. The repo is at 490 stars. The appeal is that agents defined this way stay compatible with ordinary Python testing, tracing, and refactoring.