NVIDIA open-sources NOOA, a single-class Python agent framework

A quiet framework drop from NVIDIA Labs is worth a look even if you're framework-fatigued, because it takes a swing at the thing that actually hurts about building agents right now, the sprawl across prompt files, tool schemas, callbacks and workflow graphs. Marktechpost reports that NOOA, short for NVIDIA Object-Oriented Agents, collapses all of that into a single Python class. Methods are the actions the model can take, fields are agent state, docstrings are the prompts, and type annotations are contracts the runtime enforces. Methods with a body of `...` become agentic methods completed by an LLM-driven loop, while regular methods stay deterministic Python.

The numbers NVIDIA is putting behind it are the interesting part. They report 82.2% on SWE-bench Verified with GPT-5.5, 86.8% on CyberGym L1 with network access blocked, and 85.1% mean RHAE on ARC-AGI-3 with GPT-5.6-sol. The claim that got my attention isn't the top-line accuracy though, it's the efficiency: roughly 1.1 million tokens and 28 model calls per task, against about 2.2 million tokens and 66 calls for the approaches they compared against. Half the tokens for the same task, if it survives your own runs, is real money at production scale.

Under the hood there are two execution strategies, a PredictStrategy for single typed LLM calls with local retry, and a CodeActStrategy that gives the model a Python REPL and lets it iterate until it returns a validated result. One nice touch: arguments arrive as live Python objects, and the model sees only a bounded preview with the concrete type, true length, and a head or tail sample, so a hundred-element list renders in about thirty tokens while the full variable stays in the REPL. That reads like an engineering choice made by people who watched their own token bills.

The honest caveat is that this is an alpha research preview shipped as v0.0.8 on July 30, 2026 under Apache 2.0, requiring Python 3.12 to 3.13, so treat API stability as pending. NVIDIA is also direct that the framework's AST checks and module deny-lists are defense-in-depth, not a containment boundary; the boundary is a container, a VM, or NVIDIA's own OpenShell. What the reporting doesn't give you is a head-to-head against LangGraph or AutoGen on matched tasks with a matched model, or how the harness degrades on cheaper models than GPT-5.5.

If the token-efficiency claim replicates independently, the audience that should care most is the team currently paying frontier inference bills on multi-agent scaffolding they only half understand. A single-class abstraction that reportedly does more with less is worth a weekend benchmark before the next framework commitment.