Key Takeaways
- NOOA represents agents as single Python classes, integrating capabilities, state, and prompts through methods, fields, and docstrings, with type annotations as enforced contracts.
- The six harness capabilities—typed I/O, pass-by-reference, code as action, programmable loops, explicit state, and model-callable APIs—drive superior performance and efficiency.
- Benchmark results show NOOA achieving 82.2% on SWE-bench Verified with GPT-5.5, surpassing prior state-of-the-art while using half the tokens and eliminating context compaction.
NVIDIA Unveils NOOA: A New Paradigm for AI Agent Harnesses
On July 27, 2026, NVIDIA researchers Ricardo Silveira Cabral and Paul Furgale published a technical report introducing NOOA (NVIDIA Object-Oriented Agents), an open-source framework that redefines how AI agents are constructed. By modeling an agent as a single Python class—where methods are actions, fields are state, docstrings are prompts, and type annotations are enforced contracts—NOOA transforms agent development into traditional software engineering. The result is measurable: state-of-the-art accuracy on SWE-bench Verified at 82.2% with GPT-5.5, using roughly half the tokens of competing harnesses, and without requiring context compaction.
The framework identifies six model-facing interface capabilities that collectively drive both performance and cost efficiency. This article breaks down each capability, reviews benchmark outcomes, and examines what NOOA’s approach means for the broader AI industry.
Table of Contents
The Six Capabilities Explained
Typed Input/Output
Agentic calls now support typed arguments and validated return values, moving beyond free-text exchanges. Type annotations serve as enforced contracts, ensuring that actions produce predictable outputs.
Pass by Reference
Instead of serializing tool results into the context window, NOOA allows the model to operate on live Python objects with bounded previews. This eliminates round-tripping through the context window and significantly reduces token consumption.
Code as Action
Agents act by writing and executing Python code, integrating control flow and inline method calls directly. This merges exploration and verification into a single typed environment, making the agent’s reasoning transparent and testable.
Programmable Loop Engineering
Orchestration loops are expressed as ordinary Python code, modifiable by both developers and the model itself. This flexibility allows agents to adapt their own reasoning strategies dynamically.
Explicit Object State
Durable, typed state lives on the agent object rather than within conversation history. The agent curates its own memory through model-callable tools, storing records in a human-readable SQLite file with types, importance, tags, and relationships.
Model-Callable Harness APIs
Context blocks and event history are exposed as APIs the model can inspect and manage. This allows the agent to actively curate relevant information rather than passively receiving all history.
Benchmark Dominance and Cost Efficiency
NOOA’s performance is validated across multiple domains. On SWE-bench Verified, the framework achieves 82.2% with GPT-5.5, surpassing the previous state-of-the-art of 79.2%. Critically, this accuracy comes at approximately half the token cost: NOOA uses roughly 1.1 million tokens per task, while competing harnesses require 2.2 million to reach lower scores. No context compaction is needed because pass-by-reference keeps tool results out of the context window, allowing prompt cache hits to compound.
In cybersecurity, on CyberGym L1, NOOA solves 86.8% of tasks with GPT-5.5, making it the top open-source agent on that benchmark—ahead of most closed-source systems. The agent operated with network access blocked, relying solely on reasoning over code. For ARC-AGI-3, a challenging general reasoning benchmark, a single NOOA agent reaches 50.2% mean RHAE with GPT-5.5 and 85.1% with GPT-5.6-sol, all at under $20 per game. The memory subsystem alone contributes an +11.8 point improvement over file-based notes.
The paper also reports strong results on Terminal-Bench 2.0, another agentic benchmark. These results demonstrate that identical NOOA interfaces generalize across software engineering, cybersecurity, and general reasoning tasks without domain-specific prompting.
Strategic Implications for Enterprise AI
NOOA’s open-source release has significant implications for the AI industry. By making the framework, tests, and evaluation methodology public, NVIDIA invites rigorous inspection and community contribution. This transparency is critical for enterprise adoption, where trust and reproducibility are paramount.
The efficiency gains—doubling performance per token—directly impact the economics of deploying AI agents. For organizations running large-scale agent fleets, halving token consumption without sacrificing accuracy translates to substantial cost savings. Moreover, the ability to inspect and version agents using standard software engineering tools (diff, code review, unit tests) lowers the barrier to production deployment.
From a competitive standpoint, NOOA challenges existing agent frameworks by demonstrating that harness architecture—not just model choice—is a primary driver of performance. This insight could shift how the industry prioritizes investment: toward building smarter, more efficient orchestration layers rather than solely pursuing larger models. The fact that NOOA achieves these results with a generic 253-line agent on SWE-bench, without benchmark-specific prompts, underscores the power of its core design principles.
As noted in the real-time research community coverage on HuggingFace Daily Papers, NOOA is model-agnostic and integrates seamlessly with existing Python ecosystems. This interoperability makes it a viable candidate for enterprises seeking to future-proof their agent deployments.
The Future of Agent Harness Design
NOOA represents a step forward in formalizing the interface between AI models and the infrastructure that supports them. By treating the harness as a first-class software engineering object, NVIDIA has shown that agent performance can be systematically improved through code architecture rather than brute-force prompting. The framework’s six capabilities provide a blueprint for building agents that are more efficient, more inspectable, and more reliable.
As the open-source community adopts and extends NOOA, we can expect rapid iteration on memory mechanisms, tool use, and multi-agent coordination. For teams building AI products, the lessons from NOOA are clear: invest in harness design as carefully as model selection, and treat agents as software systems rather than prompt wrappers.
For those looking to harness these advances in their own deployments, programmatic AI automation offers a pathway to integrate similar orchestration patterns into production workflows. Andres SEO Expert provides expertise in building AI pipelines that mirror the efficiency and transparency demonstrated by NOOA. To explore how these principles can be applied to your projects, connect with Andres or learn more about Andres SEO Expert‘s approach to AI-driven optimization.
Frequently Asked Questions
What is NOOA and how does it redefine AI agent construction?
NOOA (NVIDIA Object-Oriented Agents) is an open-source framework that models an AI agent as a single Python class—where methods are actions, fields are state, docstrings are prompts, and type annotations serve as enforced contracts. This transforms agent development into traditional software engineering, making agents more inspectable, testable, and efficient.
How does NOOA achieve its cost efficiency and token reduction?
NOOA uses pass-by-reference to keep tool results out of the context window, eliminating round-tripping and reducing token consumption. It requires roughly 1.1 million tokens per task on SWE-bench Verified, about half of competing harnesses (2.2 million), while achieving higher accuracy. This allows prompt cache hits to compound without needing context compaction.
What are the six model-facing interface capabilities of NOOA?
The six capabilities are: Typed Input/Output (enforced contracts via type annotations), Pass by Reference (operating on live Python objects with bounded previews), Code as Action (writing and executing Python code for control flow), Programmable Loop Engineering (orchestration loops as modifiable Python code), Explicit Object State (durable typed state on the agent object, stored in SQLite), and Model-Callable Harness APIs (context and history as inspectable APIs).
What benchmark results does NOOA achieve, and how does it compare to state-of-the-art?
On SWE-bench Verified, NOOA achieves 82.2% with GPT-5.5, surpassing the previous state-of-the-art of 79.2% at half the token cost. On CyberGym L1, it solves 86.8% of tasks with GPT-5.5, making it the top open-source agent. On ARC-AGI-3, a single NOOA agent reaches 50.2% RHAE with GPT-5.5 and 85.1% with GPT-5.6-sol at under $20 per game. The memory subsystem alone improves ARC-AGI-3 by +11.8 points.
What are the strategic implications of NOOA for enterprise AI adoption?
NOOA’s open-source release invites rigorous inspection and community contribution, which is critical for enterprise trust and reproducibility. The cost savings from halving token consumption directly impact deployment economics. By enabling agents to be inspected, versioned, and tested with standard software engineering tools (diff, code review, unit tests), NOOA lowers the barrier to production deployment and challenges the notion that only larger models drive performance.
How does NOOA differ from existing agent frameworks, and what does it imply for future harness design?
NOOA treats the harness as a first-class software engineering object rather than a prompt wrapper. Its six capabilities provide a blueprint for building more efficient, inspectable, and reliable agents. The framework demonstrates that harness architecture—not just model choice—is a primary driver of performance, shifting industry investment toward smarter orchestration layers. Its model-agnostic design and integration with Python ecosystems make it a future-proof choice for enterprises.
