Terrill Dicki
Jul 27, 2026 09:38
NVIDIA introduces NOOA, an open-source framework optimizing AI agent harnesses for higher efficiency and accuracy across benchmarks.
NVIDIA has launched the NVIDIA Labs Object-Oriented Agents (NOOA) framework, aiming to redefine how AI agents perform through optimized harness engineering. Announced on July 27, 2026, NOOA offers an open-source, Python-based architecture designed to enhance agent efficiency, accuracy, and cost-effectiveness across industries like software engineering, cybersecurity, and general reasoning.
Harness design, often overlooked, plays a pivotal role in AI performance. The NOOA framework positions the harness—the system layer surrounding AI models—as a critical factor in delivering performance gains without altering base models. NVIDIA claims that NOOA can deliver double-digit percentage improvements in benchmarks while reducing token usage and operational costs by up to 50%.
Performance Gains Across Key Metrics
NOOA has demonstrated notable results across various benchmarks. On SWE-Bench Verified, the framework achieved an 82.2% accuracy rate using GPT-5.5, surpassing the previous state-of-the-art score of 79.2%. This was done while using only 1.1 million tokens per task, compared to the 2.2 million tokens required by less efficient harnesses. Cybersecurity tests on CyberGym L1 showed similar success, with NOOA solving 86.8% of tasks—leading among open-source agents.
In general reasoning, NOOA reached 85.1% RHAE (Relative Hypothesis Accuracy Error) on the ARC-AGI-3 benchmark using GPT-5.6-sol, a significant leap from previous baselines. Notably, these results were achieved at a cost of under $20 per task, emphasizing the framework’s cost efficiency.
The Six Pillars of NOOA’s Approach
At the core of NOOA’s architecture are six capabilities designed to optimize agent performance:
Typed input/output: Ensures data consistency with validated arguments and returns.
Pass by reference: Allows the model to interact with live Python objects rather than serialized data, reducing token overhead.
Code as action: Enables models to execute Python code directly, streamlining workflows.
Programmable loop engineering: Facilitates developer-controlled iteration loops.
Explicit object state: Maintains durable, typed states for better context management.
Model-callable APIs: Provides tools for inspecting and managing context and event histories.
These features allow NOOA agents to perform more like traditional software programs, enabling extensive debugging, version control, and human-AI collaboration during development.
Cost Efficiency and Real-World Applications
NOOA’s pass-by-reference mechanism and efficient context management significantly reduce token consumption. For example, while competing harnesses use up to 2.2 million tokens for SWE-Bench tasks, NOOA achieves parity or better accuracy with half the tokens, eliminating the need for costly context compaction strategies.
Beyond efficiency, NOOA enables agents to accumulate knowledge across sessions without retraining, storing data in a human-readable SQLite file. This approach benefits industries like cybersecurity, where NOOA’s capabilities have been tested on tasks like vulnerability discovery. NVIDIA reported a task completion rate of 86.8% on CyberGym L1, with deterministic validation steps ensuring reliability.
A Broader Shift in Harness Engineering
NVIDIA’s NOOA joins a growing wave of research emphasizing the importance of harness design in AI systems. Studies from 2026, including Microsoft Research’s “Retrospective Harness Optimization” and the arXiv paper “Better Harnesses, Smaller Models,” have shown that refined harnesses can enable smaller models to rival larger ones at a fraction of the cost.
Industry interest in harness engineering is growing, with CIOs recognizing the potential for reduced latency, improved reproducibility, and lower operating expenses. NVIDIA’s open-source release of NOOA aims to accelerate these trends, providing the AI community with a transparent, extensible framework for further innovation.
What’s Next?
NOOA is available now as a research preview, with code and benchmarks accessible on NVIDIA’s GitHub. Developers can also pair the framework with NVIDIA’s OpenShell runtime for secure deployment.
As harness engineering emerges as a key performance lever in AI, NVIDIA’s NOOA framework sets a benchmark for what’s possible when optimizing the architecture around foundation models. With open collaboration encouraged, NOOA could pave the way for the next wave of cost-efficient, high-performing AI agents.
Image source: Shutterstock
Be the first to comment