NVIDIA-labs OO Agents: Native Python Object-Oriented Agents

Paul Furgale, Severin Klingler, James Nolan, Matt Staats, Gaia Di Lorenzo, Elisa Martinez Abad, Christian Schüller, Razvan Dinu, Alessio Devoto, Pascal Berard, Gal Kaplun, Elad Sarafian, Riccardo Roveri, Leon Derczynski, Ricardo Silveira Cabral

Read on arXiv

Type: paper
Venue: arXiv / NVIDIA Labs
Year: 2026
Source: arxiv
Access: free
Language: en
Added: 2026-08-14T19:50:00Z
Verified: 2026-08-14T19:50:00Z

Summary

NOOA (cite nvidia_oo_agents_2026; arXiv 2607.20709). Agent = class; methods = capabilities; fields = state; docstrings = prompts; types = contracts. Ellipsis bodies run LLM loops (Predict or CodeAct); real bodies stay deterministic Python. Six interface ideas vs 14 harnesses. Capability suite 88 tests × 10 models: 97.9% pass. SWE-bench Verified: 82.2% GPT-5.5 xhigh at ~28 calls / ~1.1M tokens vs OpenCode 78.6% / PI 78.2% at more tokens (published SOTA 79.2% at submission). Terminal-Bench 2.0 73.0% GPT-5.5 high. CyberGym L1 86.8% GPT-5.5, network blocked, top open-source. ARC-AGI-3: one 45–50-line world-model skill; GPT-5.5 50.2% RHAE, GPT-5.6-sol 85.1% at ~$13.3/game (raw Sol 13.3%); memory +11.8 vs file notes. Discord posted the NVIDIA developer blog. Research preview; execute model code in-process — sandbox with OpenShell/container.

Keywords

nooa · nvidia · agent-harness · codeact · swe-bench · arc-agi-3 · blog

Topics

agent harnesses, CodeAct, object-oriented agents

Research notes

Primary: arxiv abs 2607.20709. Discord posted https://developer.nvidia.com/blog/six-agent-harness-capabilities-for-higher-model-performance. Code https://github.com/NVIDIA-NeMo/labs-OO-Agents (GitHub SPDX Other; MarkTechPost/PyPI report Apache-2.0; pip nooa). Blog authors Cabral/Furgale; paper author list used. Harness/paper, not a hosted corpus, so no datasets_local row.