Building an AI agent today typically means juggling several separate pieces that do not share a common language: prompt templates live in one file, tool schemas in another, callback logic somewhere else, and the overall control flow gets modeled as a graph. NVIDIA's new open-source framework, NOOA, collapses all of that into a single ordinary Python class. A method's docstring becomes the instruction the model sees; the method's type annotations become a contract the runtime enforces on whatever the model returns; the class's fields become the agent's persistent state; and a method whose body is left as an ellipsis is automatically completed by an LLM at runtime, while a method with real code stays plain, deterministic Python that the model can call like any other function. The practical benefit is that agent behavior becomes something a normal software team can test, diff, refactor, and put under version control the way they would any other codebase, instead of treating it as a bespoke prompt-engineering artifact. One specific design choice, passing arguments to the model by reference as live Python objects rather than dumping their full contents into the context window, lets the agent work over very large in-memory data structures using only a small text preview, which NVIDIA credits for needing no context-compaction step at all during its benchmark runs. On SWE-bench Verified the reported result is 82.2 percent, and on a cybersecurity benchmark it hits 86.8 percent, both at roughly half the token cost of comparable open-source agent harnesses NVIDIA tested against. It is released under Apache 2.0 and installable via pip, but it is explicitly labeled alpha, and NVIDIA is direct that its safety checks are defense-in-depth rather than a real security boundary, so anyone running it against untrusted inputs still needs to wrap it in a container or VM. For developers evaluating agent frameworks, it is a notable bet that treating an agent as just a class is a cleaner abstraction than the graph- and schema-based frameworks that have dominated so far.