NVIDIA's AI headlines of the past two weeks belonged almost entirely to AVO: on August 22 the agent system posted a perfect 100 on ARC-AGI-3, clearing all 183 levels, and much of the coverage framed it as proof that "the model got smarter again." But another project the same company quietly open-sourced on July 30 — the Python agent framework NOOA (NVIDIA Object-Oriented Agents) — was skipped by most observers. Both events point to the same conclusion: the execution layer above the model is becoming the main battleground for capability gains.
What NOOA Is: An Agent Folded Into One Python Class
NOOA's full name is NVIDIA-labs Object Oriented Agents. It ships under Apache 2.0 and installs with pip install nooa. Its core design fits in one sentence: an agent is not a pile of config files and callbacks — an agent is a Python class. Fields on the class are state, ordinary methods are deterministic capabilities, docstrings are the prompts, and type annotations are the input-output contracts. Methods whose bodies contain only ... are implemented at runtime by an LLM-driven loop. The model is pluggable: through LiteLLM it supports Anthropic, OpenAI, Ollama, vLLM and other mainstream backends, so swapping the base model requires no changes to agent code.
The immediate payoff is engineering hygiene: agents can be tested, traced, refactored, and version-controlled like ordinary Python objects. For any team that has maintained the classic three-piece set of a prompt directory plus tool schemas plus callback configs, this subtraction is genuinely attractive.
The Other Side of 82.2%: Half the Token Bill
The hardest number on the scoreboard is 82.2% on SWE-bench Verified. The comparison reported by third-party coverage makes the point better: to reach that score, NOOA consumes roughly 1.1 million tokens and about 28 model calls per task, while peer frameworks typically run around 2.2 million tokens and 66 calls — roughly double on both counts. Reports from marktechpost and aiweekly corroborate this set of numbers. According to aitoolsrecap's review notes, NOOA also reports 86.8% on CyberGym L1 and 85.1% on ARC-AGI-3; these are vendor-reported figures, and the framework is still a v0.0.8 alpha research preview, so read them with a grain of salt.
Worth emphasizing: the 82.2% was achieved by NOOA paired with GPT-5.5. In other words, the same model, wrapped in a harness with a cleaner organization, improved on score and cost at the same time — which is exactly the logic of the AVO story: the model did not change; what changed is the layer wrapped around it.
Why an "Abstraction" Deserves News Coverage
The interesting thing about NOOA is not the score but the industry consensus it puts on the table: a large share of the reliability bottleneck in agents lies not in the model but in how interfaces are organized. Existing frameworks split prompts, tools, callbacks, and workflows into four separate abstractions, and every extra layer of glue adds surface area for errors. NOOA compresses them back into native Python class semantics, turning "writing an agent" back into "writing software." An early 1.9k stars on GitHub suggests real community interest in this direction.
A bucket of cold water, though: the README explicitly warns that NOOA is research software, that agents execute LLM-generated code, and that in-process AST checks and module deny-lists are defense-in-depth only — the real containment boundary must come from OS-level sandboxes such as containers or VMs. Before taking it to production, read the safety section first.
The "So What" for Readers
When making your next stack decision, instead of asking "which model is strongest," ask first "whether my harness deserves my model." The shared lesson of NOOA and AVO is that the same model, wrapped in execution frameworks of different quality, can differ by a passing grade. The framework layer is open source, auditable, and swappable — it deserves the same attention as model selection, and it may be the most underrated selection signal of the season.