NVIDIA open-sources NOOA agent framework, 82.2% on SWE-bench Verified
NVIDIA Labs open-sourced NOOA under Apache 2.0 on August 9, treating a Python class as the core agent abstraction, hitting 82.2% on SWE-bench Verified.
On August 9, NVIDIA Labs officially open-sourced the NOOA agent framework under Apache 2.0. The key innovation is structural: an agent is a Python class, not a collection of prompt templates and callback graphs. Methods whose body is empty are completed by an LLM at runtime. Methods with normal bodies are deterministic Python. Everything — state, prompts, actions, contracts — lives in one class that can be tested, traced, and version-controlled like ordinary software.
The 82.2% result on SWE-bench Verified is strong for an alpha-stage framework. CyberGym L1 reaches 86.8% and ARC-AGI-3 reaches 85.1%. NOOA is model-agnostic through LiteLLM and can be installed via pip install nooa, requiring Python 3.12 to 3.13.
On security, the containment warning in the README is critical: AST checks do not stop code that is actively trying to escape a sandbox. Always run NOOA agents in a container, VM, or NVIDIA OpenShell. The project is currently at v0.0.8 alpha.
NOOA aligns with NVIDIA's broader push into AI agent infrastructure. The company has previously launched runtime products like OpenShell targeting more controllable and observable agent deployment for enterprise customers. Open-sourcing NOOA lowers the bar for smaller teams to build agents.