NVIDIA open-sourced NOOA (NVIDIA Object-Oriented Agent), a Python framework for building AI agents, and published benchmarks showing double-digit performance gains at half the token cost of competing harnesses — a practical counterpoint to this week's debate over raw model scale.
What changed
According to MarkTechPost and NVIDIA's own technical blog, NOOA structures an entire agent as a single Python class: methods define the actions the model can take, fields hold state, docstrings become prompts, and type annotations act as contracts the runtime enforces. The framework is model-agnostic — it connects via LiteLLM to hosted APIs, Ollama, or vLLM endpoints — and ships under an Apache 2.0 license as an alpha-stage research preview (pip install nooa, version 0.0.8, released July 30).
In benchmarks published by NVIDIA, NOOA reached 82.2% on SWE-bench Verified using GPT-5.5 — above the prior state of the art, cited at around 79% among comparable harnesses — while using roughly 1.1 million tokens per task versus 2.2 million for comparable harnesses: half the cost for an equal or better result, under $20 per task according to NVIDIA. On CyberGym L1, a cybersecurity benchmark, NOOA solved 86.8% of tasks, leading among open-source agents.
NOOA is NVIDIA's core contribution to the Open Secure AI Alliance, the 37-plus company coalition — including Microsoft, Cloudflare, CrowdStrike, and Hugging Face — formed after the OpenAI-Hugging Face security incident we covered here in late July.
Why it matters
The framework includes typed, relational long-term memory stored in a human-readable SQLite database, plus a CLI, a trace viewer, and benchmark tooling — features built to make agent behavior testable, traceable, and auditable, not just fast. That's a different lever than the race toward ever-larger models: the performance gain here comes from agent architecture, not model size — reinforcing what we noted this week about ByteDance's scale race: raw parameter count is an engineering metric, not a business outcome.
The impact for Brazil
For IT and engineering teams in Brazil evaluating how to build or buy AI agent tooling, NOOA is a market signal, not a production-ready product recommendation — the package itself is alpha stage. But it surfaces a concrete evaluation criterion: a model-agnostic agent framework reduces single-vendor dependency, and built-in audit and traceability features make it easier to demonstrate governance to regulators and to the board — precisely the gap Arctera flagged here a few weeks ago, when only 19% of companies could prove their AI governance actually works in practice.
Entercast's take
NOOA is one more data point suggesting the mature phase of agentic AI adoption won't be won by whoever has the biggest model, but by whoever builds the best orchestration, memory, and audit layer on top of it — whichever model that happens to be. For Brazilian companies still at the pilot stage, the more useful question isn't "which model is strongest," but "is our agent architecture testable, traceable, and swappable between vendors without rework." It's that kind of early infrastructure decision that separates the pilot that never leaves the page from the pilot that scales.