Six Agent Harness Capabilities for Higher Model Performance

What happened

NVIDIA Labs published a technical blog post introducing NOOA, or NVIDIA Labs Object-Oriented Agents, an open-source research preview that reimagines how AI agents are built by making the harness — not the model — the primary driver of performance. Written by Ricardo Silveira Cabral and Paul Furgale and posted to the NVIDIA Technical Blog on July 27, 2026, the work demonstrates that the architecture surrounding a model can swing benchmark results by double digits while cutting token costs in half compared to other approaches, using the same underlying model throughout.

This piece breaks down what NOOA is, how its six core design ideas work in practice, what the benchmark results mean for practitioners, and which teams and builders stand to benefit most from these findings. It is aimed at engineers building or evaluating AI agents, architects choosing between harness designs, and anyone tracking the cost and performance tradeoffs of frontier model deployments.

NVIDIA Labs released NOOA as an open-source research preview that models an agent as a single Python class. The framework, its memory system, capability tests, and benchmark agents are all public, with code, data, and evaluations available so the community can reproduce and build on the results. On SWE-bench Verified, NOOA reaches 82.2% with GPT-5.5, above the published leaderboard state of the art of 79.2% at the time of submission, and 79.8% with Opus 4.6, using a general-purpose 253-line agent with no benchmark-specific prompts.

How it works

On CyberGym L1, NOOA solves 86.8% of tasks with GPT-5.5 and no network access, ranking as the top-scoring open-source agent and ahead of most leading closed-source systems. On ARC-AGI-3, a single NOOA agent with a 45-line world-modeling skill reaches 50.2% mean RHAE with GPT-5.5 and 85.1% with GPT-5.6-sol, both under $20 per game — $17.85 and $13.30 respectively — advancing the benchmark's score-cost Pareto frontier.

NOOA defines six model-facing interface ideas: typed input and output, pass by reference, code as action, programmable loop engineering, explicit object state, and model-callable harness APIs. An agent is a Python class whose methods are its capabilities, fields are its state, docstrings are its prompts, and type annotations are enforced contracts. A method whose body is an ellipsis is completed at runtime by an LLM-driven loop, while methods with a normal body run as deterministic Python. This means agents can be diffed, code-reviewed, unit-tested, and refactored using standard software tools.

Why it matters

The pass-by-reference mechanism is the key efficiency driver. Tool results become live Python variables composed directly in code rather than serialized into the context window as text. The model sees a typed, bounded preview; the full value stays alive in the execution environment. This keeps transcripts append-only and cache-valid, which is why NOOA reaches 82.2% on SWE-bench using just 29 LLM calls and approximately 1.1 million tokens per task, while comparison harnesses need 66 calls and 2.2 million tokens to reach 78.2%.

The headline finding is that harness design alone can account for double-digit benchmark swings and significant token cost differences with the same underlying model. NOOA does not require context compaction to solve SWE-bench with frontier models because median sessions peak at 22,000 to 72,000 prompt tokens against 200,000 to 400,000 context windows. By keeping tool results out of the context window through pass by reference, prefill cache hits compound across the full task and no summarization pass is needed.

The memory system adds a further dimension. NOOA's long-term memory is a SQLite file the agent curates deliberately through model-callable tools, storing typed records with importance, tags, and typed relationships such as "supports," "contradicts," and "derived-from." A reflection pass consolidates the store by merging duplicates and pruning stale records. On ARC-AGI-3, switching from file-based notes to the NOOA memory system improved RHAE by 11.8 points with the same agent, and the world-modeling skill alone added 8.5 points over a hypothesis-driven baseline.

Who is affected

Engineers building production agents today using prompt templates, tool schemas, callback code, or workflow graphs are the most directly affected group. NOOA offers an alternative where the entire agent fits in a single Python class that teams can version control, review, and test with the same practices they already use for ordinary code. Teams running on GPT-5.5, GPT-5.6-sol, or Opus 4.6 can verify the benchmarks against published scorecards, and the framework supports deploying agents under NVIDIA OpenShell, a secure runtime for sandboxed execution.

Researchers and evaluation teams working with SWE-bench Verified, ARC-AGI-3, or CyberGym L1 are also directly in scope. The CyberGym result is particularly notable for security-focused teams: NOOA solved 86.8% of tasks with no network access and a rule-based cheat check applied over every trajectory, meaning results come from reasoning over code rather than looking up known vulnerabilities. No cybersecurity-specific steering was used.

What to watch next

The technical report is available at arxiv.org/abs/2607.20709, and the framework code lives at github.com/nvidia-nemo/labs-OO-Agents alongside the benchmark agents and capability tests. Builders who want to verify the efficiency claims should check how their current harness handles context growth on long tasks and whether tool results are serialized into the conversation history. The gap NOOA closes — 66 LLM calls and 2.2 million tokens to reach 78.2% versus 29 calls and 1.1 million tokens to reach 82.2% — is large enough that measuring your own harness against similar tasks is a concrete next step before adopting any new framework.

The ARC-AGI-3 result with GPT-5.6-sol at 85.1% RHAE for $13.30 per game places NOOA on a meaningful part of the score-cost frontier, but the benchmark continues to evolve and the companion DreamTeam architecture that NOOA's single-agent version was ported from involves six specialized agents. Whether the single-agent approach holds that position as ARC-AGI-3 difficulty increases, and whether the same six harness capabilities transfer to multimodal or multi-agent deployments, are the open questions most worth tracking in the months ahead.