Building an AI Agent Is Now One Python Class

Building AI agents — software that plans and carries out multi-step tasks on its own — has meant learning heavyweight frameworks, writing separate config files, and wiring together a lot of boilerplate. NVIDIA just released NOOA, an open-source library that replaces all of that with a single Python class. An agent is a Python class: methods are actions the AI can take, docstrings become the instructions the model reads, and fields hold the agent's memory.

That is not just cleaner code — it also cuts costs. Agents built with NOOA use roughly half as many tokens as comparable open frameworks — about 1.1 million per task instead of 2.2 million — which maps directly to lower API bills. They make around 28 model calls per task, compared to 66 for similar tools. On SWE-bench Verified, the main benchmark for coding agents, NOOA scored 82.2%, ahead of most open alternatives.

The key trick is how NOOA handles data. Most frameworks dump large data structures into the model's context window — that is expensive and can confuse the model. NOOA keeps objects alive in a Python environment and only shows the model a short summary — about 30 tokens per object. The model acts on real, live data without being overwhelmed by it. Two execution modes are available: a single structured call for simpler tasks, and an iterative loop where the model writes and runs code until the job is done.

Any team building AI-powered software should test NOOA today. It is Apache 2.0 licensed and installs with one command: pip install nooa. One important caveat: it is still an alpha research preview, and its code-execution feature is a security risk without proper isolation. Test it inside a container or VM — never directly on a production server. For prototyping, internal tools, and coding workflows, this is one of the most practical starting points available right now.