Six Agent Harness Capabilities for Higher Model Performance

The Architecture is the Agent

NVIDIA Labs has released NOOA (NVIDIA Labs Object-Oriented Agents), an open-source framework that demonstrates how an agent's surrounding architecture, or “harness,” can be as influential as the underlying model for achieving top-tier performance. The research preview, which includes the framework, memory system, and benchmark agents, has already posted state-of-the-art results on several industry benchmarks. This release shifts focus from a model-centric view to a more holistic approach, where context rendering, state management, and action execution are treated as first-class engineering problems that can yield double-digit swings in accuracy and cost-efficiency.

Technical Framework and Capabilities

At its core, NOOA simplifies agent development by structuring an agent as a single Python class where methods are capabilities, fields are state, and docstrings serve as prompts. This object-oriented design makes agents easier to test, version, and review using standard software development tools. The framework's performance is driven by six key capabilities that shape how the model interacts with its environment:

• Typed input/output for validated arguments and returns.
• Pass-by-reference operations on live Python objects instead of serialized text.
• Code-as-action, allowing the model to act by writing Python.
• Programmable orchestration loops written in standard Python.
• Explicit, durable object state living on the agent.
• Model-callable APIs for inspecting context and event history.

Additionally, NOOA includes a long-term memory subsystem where agents actively curate a knowledge graph in a human-readable SQLite store. This approach improved performance on the ARC-AGI-3 benchmark by +11.8 points compared to using simple file-based notes.

Impact on Efficiency and the Open Ecosystem

The architectural choices in NOOA lead to substantial efficiency gains. On the SWE-bench Verified benchmark, the framework achieved 82.2% accuracy with GPT-5.5 using roughly half the LLM calls and tokens (~1.1M) that comparison harnesses needed to reach lower scores (78.2% at 2.2M tokens). This is primarily due to the pass-by-reference mechanism, which avoids bloating the context window and eliminates the need for context compaction or summarization. By making the framework, tests, and evaluation methodology public, NVIDIA is providing the AI community with a reproducible and inspectable system that highlights the critical role of harness engineering in building more capable and cost-effective agents.

The performance gains from NOOA underscore a critical industry insight: advancements in agentic AI will increasingly depend on sophisticated harness engineering—how context, state, and actions are managed—rather than solely on scaling the underlying language models.