Terrill Dicki
Jul 27, 2026 09:38
NVIDIA introduces NOOA, an open-source framework optimizing AI agent harnesses for greater effectivity and accuracy throughout benchmarks.
NVIDIA has launched the NVIDIA Labs Object-Oriented Brokers (NOOA) framework, aiming to redefine how AI brokers carry out via optimized harness engineering. Introduced on July 27, 2026, NOOA presents an open-source, Python-based structure designed to boost agent effectivity, accuracy, and cost-effectiveness throughout industries like software program engineering, cybersecurity, and basic reasoning.
Harness design, usually missed, performs a pivotal position in AI efficiency. The NOOA framework positions the harness—the system layer surrounding AI fashions—as a important consider delivering efficiency good points with out altering base fashions. NVIDIA claims that NOOA can ship double-digit proportion enhancements in benchmarks whereas lowering token utilization and operational prices by as much as 50%.
Efficiency Good points Throughout Key Metrics
NOOA has demonstrated notable outcomes throughout varied benchmarks. On SWE-Bench Verified, the framework achieved an 82.2% accuracy charge utilizing GPT-5.5, surpassing the earlier state-of-the-art rating of 79.2%. This was carried out whereas utilizing only one.1 million tokens per activity, in comparison with the two.2 million tokens required by much less environment friendly harnesses. Cybersecurity assessments on CyberGym L1 confirmed related success, with NOOA fixing 86.8% of duties—main amongst open-source brokers.
Normally reasoning, NOOA reached 85.1% RHAE (Relative Speculation Accuracy Error) on the ARC-AGI-3 benchmark utilizing GPT-5.6-sol, a major leap from earlier baselines. Notably, these outcomes have been achieved at a value of below $20 per activity, emphasizing the framework’s price effectivity.
The Six Pillars of NOOA’s Strategy
On the core of NOOA’s structure are six capabilities designed to optimize agent efficiency:
- Typed enter/output: Ensures knowledge consistency with validated arguments and returns.
- Go by reference: Permits the mannequin to work together with dwell Python objects somewhat than serialized knowledge, lowering token overhead.
- Code as motion: Permits fashions to execute Python code immediately, streamlining workflows.
- Programmable loop engineering: Facilitates developer-controlled iteration loops.
- Specific object state: Maintains sturdy, typed states for higher context administration.
- Mannequin-callable APIs: Gives instruments for inspecting and managing context and occasion histories.
These options permit NOOA brokers to carry out extra like conventional software program packages, enabling in depth debugging, model management, and human-AI collaboration throughout growth.
Value Effectivity and Actual-World Functions
NOOA’s pass-by-reference mechanism and environment friendly context administration considerably cut back token consumption. For instance, whereas competing harnesses use as much as 2.2 million tokens for SWE-Bench duties, NOOA achieves parity or higher accuracy with half the tokens, eliminating the necessity for expensive context compaction methods.
Past effectivity, NOOA allows brokers to build up information throughout periods with out retraining, storing knowledge in a human-readable SQLite file. This strategy advantages industries like cybersecurity, the place NOOA’s capabilities have been examined on duties like vulnerability discovery. NVIDIA reported a activity completion charge of 86.8% on CyberGym L1, with deterministic validation steps making certain reliability.
A Broader Shift in Harness Engineering
NVIDIA’s NOOA joins a rising wave of analysis emphasizing the significance of harness design in AI techniques. Research from 2026, together with Microsoft Analysis’s “Retrospective Harness Optimization” and the arXiv paper “Higher Harnesses, Smaller Fashions,” have proven that refined harnesses can allow smaller fashions to rival bigger ones at a fraction of the price.
Trade curiosity in harness engineering is rising, with CIOs recognizing the potential for decreased latency, improved reproducibility, and decrease working bills. NVIDIA’s open-source launch of NOOA goals to speed up these traits, offering the AI group with a clear, extensible framework for additional innovation.
What’s Subsequent?
NOOA is obtainable now as a analysis preview, with code and benchmarks accessible on NVIDIA’s GitHub. Builders may pair the framework with NVIDIA’s OpenShell runtime for safe deployment.
As harness engineering emerges as a key efficiency lever in AI, NVIDIA’s NOOA framework units a benchmark for what’s attainable when optimizing the structure round basis fashions. With open collaboration inspired, NOOA may pave the way in which for the subsequent wave of cost-efficient, high-performing AI brokers.
Picture supply: Shutterstock