NVIDIA NOOA vs LangChain: Python agents and token costs

NVIDIA NOOA (labs-OO-Agents) is the subject of this review of Python-based agent design. The speaker explains how docstrings become prompts, type annotations define contracts and object fields hold state. Methods with an ellipsis body let the model write and execute code at runtime, while implemented methods remain ordinary Python. Models can work with object references and use loops rather than request each operation separately.

The local AI connection is backend flexibility. According to the speaker, NOOA uses LiteLLM to support hosted providers and models on the user's computer, including through vLLM without an API key. Python 3.12 or later is required. The discussion compares its model-visible capabilities with LangChain, LangGraph and other frameworks rather than demonstrating an installation.

The speaker reports NVIDIA's SWE-bench result of 82.2%, using about 28 model calls and 1.1 million tokens per task, against a competing harness using 66 calls and 2.2 million tokens. These are NVIDIA's own evaluations; the speaker says outside teams had not reproduced them and notes differing evaluation budgets for ARC-AGI-3.

NOOA is presented as open source research software, with package version 0.0.8 and an API subject to change. The security discussion stresses that generated-code checks are insufficient isolation: execution needs a container or virtual machine. Type checks also cannot establish whether an agent's business decision is correct.