NVIDIA Open-Sources NOOA, Unifying Agent State, Tools, and Persistent Memory in a Python Class

NVIDIA Labs released a research preview of Object-Oriented Agents (NOOA) on July 27, aiming to bring agent frameworks back into conventional Python software engineering. Developers describe an agent with a single class: fields store state, methods represent capabilities, docstrings provide prompts, and type annotations serve as input/output contracts. If a method body contains an ellipsis, it is completed at runtime by an LLM loop; methods with ordinary function bodies remain deterministic Python. This design allows agents to work directly with diffs, unit tests, tracing, and version control, without separately maintaining tool schemas, workflow graphs, and prompt templates.

NOOA's key innovation goes beyond syntax. The model writes Python through a Jupyter-style REPL, retaining tool results as live objects in the execution environment while showing only bounded previews in the context. Compared with repeatedly serializing large JSON payloads or terminal output back into the prompt, pass-by-reference can shorten the transcript while preserving append-only prompt caching. Long-term memory is stored in a readable SQLite file. Records include types, importance scores, and tags, and can form a small knowledge graph through relationships such as `supports`, `contradicts`, and `derived-from`. Agents can proactively write, correct, and query records, while background processes merge duplicates and remove stale content.

NVIDIA reports that, using the same GPT-5.5 base model, a 253-line general-purpose agent achieved 82.2% on SWE-bench Verified, averaging about 29 model calls and 1.1 million tokens per task. A comparison system required 66 calls and 2.2 million tokens to reach 78.2%. NOOA also reportedly achieved 86.8% on CyberGym L1 and, paired with GPT-5.6 Sol, obtained 85.1% mean RHAE on ARC-AGI-3 at a cost of approximately $13.30 per game. Code, data, and evaluation procedures accompany these results, but the comparison points were drawn from public results available at submission time; the models, budgets, and harnesses were not fully aligned. The gains therefore cannot yet be attributed entirely to the object-oriented interface.

The library currently must be installed directly from GitHub, while the CLI, memory features, and evaluation pipeline remain optional subpackages. More importantly, “code as action” means the model can execute arbitrary Python. The project explicitly warns that this may exfiltrate data, delete files, or modify the environment, and recommends using an isolated execution environment such as OpenShell. Engineering teams should next verify whether type contracts remain stable during error retries, how SQLite-based memory handles sensitive data, and whether the token-efficiency gains can be reproduced across models and real-world enterprise toolchains.