Details
- NVIDIA AI announces it is contributing open models, weights, data, and research to the Open Secure AI Alliance.
- As part of this contribution, NVIDIA is releasing NVIDIA Labs Object-Oriented Agents (NOOA) as open agent harness research.
- NOOA is a native Python, object-oriented framework for building AI agents, where agents are structured as classes with methods, state fields, and type annotations for contracts.
- The framework focuses on harness engineering: memory systems, capability tests, and benchmark agents, enabling reproducible evaluation and iteration by the broader community.
- According to NVIDIA’s research preview, NOOA reaches state-of-the-art results across software engineering, cybersecurity, and complex reasoning benchmarks compared with prior open-source agent frameworks.
- Benchmarks such as SWE-Bench Verified and Terminal-Bench 2.0 show NOOA-based agents achieving higher accuracy with fewer tokens than baselines that rely on heavier prompt orchestration or DSLs.
- NVIDIA positions NOOA as an open, extensible foundation others can challenge and build on, while aligning with the Alliance’s focus on secure, transparent AI tooling.
- The announcement also highlights collaboration with Alliance participants, including researchers such as @zanoga, signaling joint work on open, secure agent infrastructure.
- Official materials describe NOOA as a research preview that is available with code, data, and evaluations, supporting community-led verification and improvement.
Impact
By contributing NOOA and other open assets to the Open Secure AI Alliance, NVIDIA strengthens its role in shaping open, secure AI infrastructure while keeping pace with rivals pushing agent frameworks into the mainstream. Open-sourcing a state-of-the-art agent harness could accelerate best practices around safety, evaluation, and efficiency, and may pressure other major AI labs to share more of their agent tooling rather than only models.
