|

NVIDIA AI Releases NOOA: An Object-Oriented Python Framework That Turns an AI Agent Into a Single Python Class

ByRicardo

NVIDIA Labs has open-sourced NOOA (NVIDIA Object-Oriented Agents), a model-agnostic Python framework for constructing AI brokers. Agent improvement in the present day is break up throughout immediate templates, software schemas, callback code, and workflow graphs. NOOA collapses all of it into one Python class. Methods are the actions the mannequin can take. Fields are agent state. Docstrings are prompts. Type annotations are contracts the runtime enforces. A way whose physique is ... is accomplished at runtime by an LLM-driven loop, whereas a methodology with a regular physique stays deterministic Python. Developers and fashions subsequently share one interface, so agent habits could be examined, traced, refactored, and version-controlled like unusual software program. NVIDIA studies 82.2% on SWE-bench Verified, 86.8% on CyberGym L1, and 85.1% imply RHAE on ARC-AGI-3 — at roughly half the tokens of the open harnesses it was in contrast towards.

Is it deployable?

Yes, however solely inside OS-level isolation. NOOA is Apache 2.0, installs with pip set up nooa (v0.0.8, launched July 30, 2026), and requires Python 3.12–3.13. PyPI classifies it as alpha, and NVIDIA describes it as a analysis preview. Agents can execute LLM-generated code, and NVIDIA states instantly that its AST checks and module deny-lists are defense-in-depth guardrails, not a containment boundary. The containment boundary is a container, a VM, or NVIDIA OpenShell. Models are pluggable via LiteLLM, so hosted APIs, Ollama, and vLLM endpoints all work.

  • Company stage: AI-native startups and mid-market platform groups constructing inner brokers. Enterprise AI platform and applied-research teams working evaluations or pilots. Regulated manufacturing workloads ought to await a secure launch.
  • Industries: developer tooling, cybersecurity, cloud and DevOps, knowledge analytics, monetary providers operations, buyer help.
  • Applications: repository concern triage and patching, terminal and infrastructure automation, vulnerability validation pipelines, large-batch classification and extraction over in-memory knowledge, typed multi-agent orchestration.

An agent is a Python object

NVIDIA Labs launched NOOA (NVIDIA Object-Oriented Agents), a model-agnostic Python framework for constructing brokers. Traditional agent improvement splits supply throughout immediate templates, software schemas, callbacks, and workflow graphs. NOOA collapses that into one class.

Methods are the actions the mannequin can take. Fields are state. Docstrings are prompts. Type annotations are contracts enforced by the runtime. A way whose physique is ... turns into an agentic methodology, accomplished at runtime by an LLM-driven loop; a methodology with a regular physique stays deterministic Python the mannequin can name as a software.

Two methods ship. PredictTechnique is a single typed LLM name with a native retry loop on validation failure. CodeActStrategy runs an iterative Python REPL the place the mannequin calls execute_python(...) till it submits return_result(...), which is validated towards the return annotation.

Six capabilities on one floor

The research team identifies six model-facing concepts it claims to be the primary to mix: typed enter/output, go by reference over reside objects, code as motion, programmable loop engineering, express object state, and model-callable harness APIs. NVIDIA scored fourteen frameworks and harnesses—LangGraph, Google ADK, PydanticAI, smolagents, Claude Agent SDK, OpenAI Codex, OpenHands, and others—towards the identical axes, and studies partial protection in all places else.

Pass by reference is the load-bearing one. Arguments arrive as reside Python objects; the mannequin sees solely a bounded preview with the concrete kind, true size, and a head/tail pattern. 100-element checklist renders in about thirty tokens whereas the total variable stays within the REPL. Context is break up into a cacheable static prefix, an append-only typed occasion historical past, and dynamic blocks on the tail, which preserves KV-cache reuse throughout turns.

An non-compulsory reminiscence subsystem attaches to an unmodified agent. Seven model-callable instruments write and recall information ranked by ACT-R activation, multi function human-inspectable SQLite file.

Performance

Capability assessments ran 88 assessments 5 occasions throughout ten fashions: 4,309 of 4,400 information handed (97.9%). A six-family stress subset overlaying batching, error restoration, and decomposition handed 84.7%, the place the hole between small and frontier fashions widens from 3.2 to 23 factors.

End-to-end, a benchmark-agnostic 253-line agent reaches 82.2% on SWE-bench Verified with GPT-5.5 at xhigh effort, towards 78.6% for OpenCode and 78.2% for PI, and 79.8% with Opus 4.6. On Terminal-Bench 2.0 it reaches 73.0% at excessive effort versus 60.7% and 68.5%, although PI leads at xhigh with 75.3%. On CyberGym L1 it solves 86.8% with community entry blocked, the highest open-source outcome reported. On ARC-AGI-3, one agent with a one-page world-model talent reaches 50.2% imply RHAE with GPT-5.5 and 85.1% with GPT-5.6-sol, underneath $20 per sport.

Efficiency is the extra attention-grabbing outcome: 82.2% at roughly 1.1M tokens and ~28 mannequin calls per process, towards 2.2M tokens and 66 requires PI at 78.2%. Trace evaluation additionally credit validated termination—OpenCode stops when the mannequin replies with out a software name, whereas NOOA requires a typed TaskResult carrying proof and a verification command.