NVIDIA open-sources NOOA: an agent is one Python class, 82% SWE-bench
A model-agnostic harness, pluggable via LiteLLM to any API, Ollama or vLLM, that runs agents on ~half the tokens and tops the open-source CyberGym board.
Copy markdown
An agent is one Python class
Methods are the actions, fields are the state, docstrings are the prompts, and type hints become runtime contracts. A method with a `...` body runs as an LLM loop; ordinary methods stay deterministic tools the model can call. Install is `pip install nooa` on Python 3.12–3.13.
82.2% SWE-bench on half the tokens
NOOA posts 82.2% on SWE-bench Verified (GPT-5.5) and 86.8% on CyberGym L1 — the top-scoring open-source agent — while spending about 1.1M tokens and 28 model calls per task, versus 2.2M tokens and 66 calls for comparable harnesses.
Pass-by-reference keeps big data out of context
Large objects stay live in the Python REPL while the model sees only bounded previews — roughly 30 tokens for a 100-element list. That, not a bigger context window, is where the token savings come from.
Bring your own model via LiteLLM
Models are pluggable through LiteLLM, so hosted APIs, local Ollama, and vLLM endpoints all drop in. You can wire an object-oriented agent onto a local model this weekend without rewriting the harness.
Alpha — sandbox before you ship
It's a research preview (v0.0.8). AST checks and module deny-lists are defense-in-depth, not a containment boundary, so LLM-written code needs real OS isolation — a container, VM, or NVIDIA OpenShell. Keep it out of production for now.