Editorial illustration for Six-Agent DreamTeam Architecture Coordinates for Higher Model Performance
Six-Agent Architecture Boosts Model Performance
NVIDIA Labs released an open-source research preview called NOOA, short for NVIDIA Labs Object-Oriented Agents, built around a finding that gets buried in most model comparisons: the harness matters as much as the model. Swap out the scaffolding around an LLM, the code that renders context, fires off tool calls, tracks state, and decides when a task is finished, and benchmark scores can shift by double digits even with identical weights underneath. Token costs move too, sometimes by a lot.
NOOA packages this idea into a Python framework, complete with a memory system, capability tests, and benchmark agents. NVIDIA published the code, the data, and the evaluation results, betting that other researchers will want to poke holes in the approach or extend it. The framework's core move is structural: it collapses the usual sprawl of prompt templates, tool schemas, and callback logic into something closer to a single object.
That object is a Python class. Its methods double as capabilities. Its fields hold state. Even its docstrings do work, functioning as prompts read by the model at runtime.
The harness is the architecture surrounding the model. How it renders context, executes actions, manages state, and decides when a task is done shapes outcomes just as much as the model itself. Harness design alone can account for double-digit swings in benchmark results and significant differences in token cost, with the same underlying model.
Why this matters
NOOA's real contribution isn't the six-agent DreamTeam setup itself, it's the proof that harness design carries as much weight as model choice. NVIDIA Labs found double-digit swings in benchmark results from the same underlying model, just by changing how context gets rendered and actions get executed. For developers, that's a direct challenge to the current obsession with model selection: you can burn weeks fine-tuning or swapping models and still lose to a competitor who just built a better skill file.
The single-agent port, with one 45-line skill replacing six coordinating agents, is the more interesting data point here. If a stripped-down harness can approximate a six-agent system's performance, that says something about where the actual complexity in "agentic AI" lives, and it's not always where teams assume. Worth watching: whether NOOA's open-source release lets outside teams reproduce those token-cost differences on their own benchmarks, or whether this turns out to be another result that looks great in NVIDIA's own test harness and shrinks under independent scrutiny.
Related Reading
Research & Benchmarks Google's FACTS benchmark shows 70% factuality ceiling across four tests
Research & Benchmarks Databricks finds multi-step agents beat single-turn RAG by 21% to 38% on STaRK
Research & Benchmarks Nvidia's DLSS 4.5 beta adds 6x Multi Frame Generation for RTX 50 GPUs
From Archives Cursor Claims Kimi K2.5 Model Shows Cheaper AI Can Code With Frontier Model Planning
From Archives OpenAI Models Escaped Containment for Days in Hugging Face Breach
Common Questions Answered
What is NOOA and what does it stand for according to NVIDIA Labs?
NOOA stands for NVIDIA Labs Object-Oriented Agents and is an open-source research preview released by NVIDIA Labs. It was built around the finding that the harness architecture surrounding an LLM matters as much as the model itself in determining performance outcomes.
How much can benchmark scores change by modifying the harness around an LLM?
According to NVIDIA Labs research, benchmark scores can shift by double digits even when using identical model weights underneath. The harness design alone—including how it renders context, executes actions, manages state, and decides task completion—can account for these significant performance variations.
What specific components of the harness architecture impact model performance?
The harness impacts performance through four key mechanisms: how it renders context, how it fires off and executes tool calls, how it tracks and manages state, and how it decides when a task is finished. These architectural decisions can produce double-digit swings in benchmark results and significant differences in token costs with the same underlying model.
Why should developers reconsider their focus on model selection according to this research?
NVIDIA Labs found that harness design carries as much weight as model choice in determining outcomes, meaning developers can achieve better results by optimizing the scaffolding architecture rather than spending weeks fine-tuning or swapping between different models. This challenges the current industry obsession with model selection as the primary performance lever.
What impact does harness design have on token costs?
Token costs can move significantly based on harness design changes, sometimes by a substantial amount, even when the underlying model remains identical. This means that optimizing the harness architecture can provide cost efficiency benefits alongside performance improvements.
Further Reading
- Generative Multi-Agent World Modeling Beyond Two Players - NVIDIA Research
- How to Build Deep Agents for Enterprise Search with NVIDIA AI-Q and LangChain - NVIDIA Developer Blog
- Build Specialized AI Agents: Post-GTC Developer Deep Dive - NVIDIA Developer
- LangChain Announces Enterprise Agentic AI Platform Built ... - LangChain
- AI Startup Solutions | NVIDIA Innovation Lab - NVIDIA
