NVIDIA Labs has open-sourced NOOA (NVIDIA Object-Oriented Agents), a model-agnostic Python framework that consolidates AI agent development into a single class. Released July 30 as version 0.0.8, the project eliminates the need to manage scattered prompt templates, tool schemas, and workflow graphs. Developers build agents by defining methods for model actions, fields for state, docstrings for instructions, and type annotations for runtime contracts.
Architecture and execution model
The framework ships two execution strategies. PredictStrategy routes requests through a single typed LLM call with a local retry loop when validation fails. CodeActStrategy runs an iterative Python REPL where the model executes commands until it submits a return statement validated against the function annotation.
Building structured AI Agents & Automation often requires separating state from logic, which NOOA handles through explicit Python fields and methods. In its documentation, NVIDIA wrote that "A method whose body is ... is completed at runtime by an LLM-driven loop, while a method with a normal body stays deterministic Python the model can call as a tool."
Capability mechanics and context handling
Pass-by-reference remains the core mechanic. Arguments arrive as live Python objects, but the model only sees a bounded preview containing the concrete type, actual length, and a head-tail sample. This design keeps large datasets in the REPL without forcing context compaction.
The system splits context into a cacheable static prefix, an append-only event history, and dynamic tail blocks to preserve KV-cache reuse across turns. An optional memory subsystem writes records to a human-inspectable SQLite file using ACT-R activation ranking.
Benchmark results and token efficiency
Capability tests ran 88 scenarios five times across ten models, passing 4,309 of 4,400 records. End-to-end benchmarks show the framework scoring 82.2% on SWE-bench Verified and 86.8% on CyberGym L1. These results came at roughly half the token cost of competing open harnesses, averaging 1.1M tokens per task instead of 2.2M.
Trace analysis credits this efficiency to validated termination protocols that require typed result objects carrying evidence and verification commands. The complete repository is available on the GitHub repository.
Deployment constraints and safety
NVIDIA classifies NOOA as an alpha research preview requiring Python 3.12 or 3.13. The package installs via pip under an Apache 2.0 license and plugs into existing Generative AI and LLM infrastructure through LiteLLM, supporting hosted APIs, Ollama, and vLLM endpoints.
Generated code execution demands strict environment control. The company explicitly noted that "AST checks and module deny-lists are defense-in-depth guardrails, not a containment boundary." Teams must run agents inside containers, virtual machines, or NVIDIA OpenShell to prevent uncontrolled system access.
Why this matters for IT and development teams
NOOA forces agent behavior into standard software engineering practices. Developers can test, trace, and version-control agent workflows using familiar Python tooling instead of learning proprietary graph DSLs. The token savings and validated termination protocols reduce inference costs and improve reliability in production pipelines. Teams running internal automation or evaluation pilots should treat the current release as a research preview and isolate all generated code execution. Stable versions will likely arrive once the framework proves itself outside controlled benchmark environments.
Your membership also unlocks:
