NVIDIA’s NOOA slashes agent token use in half while hitting 82.2% on SWE-bench

NVIDIA Labs just open-sourced NOOA, a Python framework that treats AI agents as… well, just Python objects. It’s an Apache 2.0 research preview, model-agnostic, and installs with a simple pip command. The pitch is straightforward: stop scattering agent logic across prompt templates, tool schemas, and workflow graphs. Collapse it all into one class where methods are actions, fields are state, and docstrings serve as the prompts. A method body that’s literally ... becomes agentic, handed off to an LLM-driven loop at runtime. Keep the body intact and it’s deterministic Python the model can call as a tool. Developers and models share one interface, which means agent behavior can be tested, traced, and version-controlled like regular software.

Two strategies ship in the box. PredictStrategy fires off a single typed LLM call with a local retry loop when validation fails. CodeActStrategy is more interesting — it’s an iterative Python REPL where the model calls execute_python(...) until it finally submits return_result(...), which then gets validated against the return annotation. The real load-bearing feature, though, is pass by reference. Arguments arrive as live Python objects, but the model only sees a bounded preview with the concrete type, true length, and a head/tail sample. NVIDIA says a hundred-element list renders in about thirty tokens while the full variable stays live in the REPL. Context is split into a cacheable static prefix, an append-only typed event history, and dynamic tail blocks — a design that preserves KV-cache reuse across turns.

The numbers are genuinely strong. A benchmark-agnostic 253-line agent hits 82.2% on SWE-bench Verified with GPT-5.5 at xhigh effort, compared to 78.6% for OpenCode and 78.2% for PI. On CyberGym L1 it solves 86.8% with network access blocked, which NVIDIA calls the top open-source result reported. ARC-AGI-3 reaches 85.1% mean RHAE with GPT-5.6-sol at under $20 per game. But the detail that should make engineering leads pay attention is efficiency: that 82.2% SWE-bench score came at roughly 1.1 million tokens and about 28 model calls — roughly half the tokens of the open harnesses it was compared against. In a world where API bills are the silent killer of agent deployment, that’s not a footnote. That’s the headline.

There’s a sharp edge here, though, and NVIDIA doesn’t hide it. The framework can execute LLM-generated code, and the company is explicit that its AST checks and module deny-lists are defense-in-depth guardrails — not a containment boundary. If you’re running this, you need a container, a VM, or NVIDIA OpenShell. PyPI classifies NOOA as alpha software, and the v0.0.8 release notes call it a research preview. For AI-native startups and platform teams building internal agents, this is immediately interesting. For anyone in a regulated production environment, you’re waiting for a stable release. The tension between capability and safety is baked right into the architecture, and no amount of clever type annotations resolves it.