NOOA Deep Dive: NVIDIA’s Pythonic AI Agents Framework with Practical Implementations

NVIDIA has unveiled NOOA (NVIDIA Object-Oriented Agents), an open-source framework that models AI agents as single Python classes, unifying capabilities, state, prompts, and memory. The framework's design, which includes typed I/O, pass-by-reference, and code-as-action, aims to reduce fragmentation in agent development. Benchmarks show NOOA achieving 82.2% on SWE-bench Verified with about 50% fewer token calls compared to traditional harnesses, and NVIDIA Labs states, 'NOOA is to AI agents what PyTorch was to deep learning: a simple interface for complex systems.'

Author: TrixSec In July 2026, NVIDIA unveiled NOOA NVIDIA Object-Oriented Agents , an open-source framework that redefines AI agents as single Python classes . By unifying capabilities, state, prompts, and memory into a cohesive interface, NOOA addresses the fragmentation in agent development while delivering performance, inspectability, and security . This guide explores NOOA’s architecture, benchmarks, and practical implementations —including a full code walkthrough of a cybersecurity agent. NOOA pronounced "no-ah" treats AI agents as Python objects , eliminating the need for: Instead, an agent is a single class where: ... pdb or pytest like normal Python. “NOOA is to AI agents what PyTorch was to deep learning: a simple interface for complex systems.”—NVIDIA Labs NOOA’s design centers on six model-facing interfaces : | Capability | Description | Example | |---|---|---| Typed I/O | Methods enforce input/output types no free text . | def scan port host: str, port: int - dict: | Pass by Reference | Agents manipulate live Python objects e.g., self.state . | self.vulnerabilities.append issue | Code as Action | Agents execute Python e.g., import socket . | socket.connect host, port | Programmable Loops | Orchestration uses standard Python for , while . | for ip in subnet: self.scan ip | Explicit Object State | State persists as fields not just in conversation history . | self.last scan = datetime.now | Harness APIs | Context/memory are Python APIs e.g., self.memory.query . | matches = self.memory.search tags= "exploit" | python from nooa import Agent from typing import Dict, List from datetime import datetime class SecurityAgent Agent : """A cybersecurity assistant for vulnerability scanning.""" def init self : self.scanned hosts: List str = Persistent state self.last scan: datetime = None def scan host self, host: str - Dict str, str : """ Scan a host for open ports and vulnerabilities. Args: host str : Target hostname/IP. Returns: Dict str, str : Report with findings. """ ... LLM implements this at runtime def add to history self, host: str - None: """Record a scanned host deterministically.""" self.scanned hosts.append host self.last scan = datetime.now NOOA’s memory subsystem stores typed, relational knowledge in a SQLite database. Key features: Each memory has: content str : The knowledge e.g., "CVE-2026-1234 affects OpenSSH 9.0" . tags List str : Categorization e.g., "vulnerability", "critical" . importance float : Priority 0.0–1.0 . relationships : Links to other records e.g., "supports" , "contradicts" .Relevant memories surface into the agent’s context during execution. Multiple agents can access the same store with separate ownership. Add a vulnerability to memory self.memory.add content="CVE-2026-1234: RCE in OpenSSH 9.0. Patch immediately.", tags= "cve", "critical", "openssh" , importance=0.9, relationships={"affects": "openssh-9.0" } Query memories later critical cves = self.memory.query tags= "cve", "critical" , limit=5 A background process: NOOA’s July 2026 benchmarks show efficiency gains over traditional frameworks: | Benchmark | NOOA GPT-5.5 | Comparison Harnesses | Token Savings | |---|---|---|---| SWE-bench Verified | 82.2% 29 calls | 78.2% 66 calls | ~50% | CyberGym L1 | 86.8% | N/A | N/A | ARC-AGI-3 | 50.2% RHAE | Baseline: ~40% | ~20% | ... methods reduce round-trips. “Harness design alone can account for double-digit swings in benchmark results—with the same underlying model.”—NVIDIA ... methods. os.system . Run agent in OpenShell container docker run -it --rm nvcr.io/nvidia/openshell:latest nooa run agent.py import subprocess . python from nooa.sandbox import DENY LIST DENY LIST.extend "subprocess", "socket", "os.system" “NOOA’s centralized design makes audits easier—but also concentrates risk. Sandboxing isn’t optional.”—Karthik Karunanithi, IBM Let’s build a vulnerability scanner agent with NOOA. python from nooa import Agent from typing import Dict, List, Optional import requests class VulnScannerAgent Agent : """Scans hosts for CVEs and suggests patches.""" def init self : self.scanned hosts: List str = self.api key: str = "" For vulnerability DBs def set api key self, key: str - None: """Securely set the API key.""" self.api key = key In production, use a secrets manager def scan host self, host: str - Dict str, List Dict : """ Scan a host for CVEs. Args: host str : Target e.g., "192.168.1.1" . Returns: Dict str, List Dict : {"vulnerabilities": ... , "suggestions": ... } """ ... LLM implements scan logic def query cve db self, cve id: str - Optional Dict : """Fetch CVE details from a database.""" headers = {"Authorization": f"Bearer {self.api key}"} response = requests.get f"https://api.cvedb.com/v1/cves/{cve id}", headers=headers return response.json if response.ok else None php def record finding self, host: str, cve: Dict - None: """Store a vulnerability in memory.""" self.memory.add content=f"{host} affected by {cve 'id' }: {cve 'description' }", tags= "vulnerability", "unpatched", host , importance=0.9, relationships={"affects": host , "type": cve "id" } def get patch suggestions self, cve id: str - List str : """Retrieve patch suggestions from memory.""" results = self.memory.query tags= "patch", cve id , limit=3 return r "content" for r in results php def full scan self, hosts: List str - Dict str, Dict : """Scan multiple hosts and aggregate results.""" report = {} for host in hosts: report host = self.scan host host for vuln in report host "vulnerabilities" : self.record finding host, vuln return report Initialize scanner = VulnScannerAgent scanner.set api key "your api key here" Scan and record results = scanner.full scan "192.168.1.1", "192.168.1.2" print results Query memory later print scanner.get patch suggestions "CVE-2026-1234" scan host LLM-driven + query cve db deterministic . List str for hosts .| Feature | NOOA | LangGraph | AutoGen | CrewAI | |---|---|---|---|---| Language | Python | Python | Python | Python | State Management | Python fields | JSON/YAML | Dicts/files | JSON | Tool Definition | Python methods | JSON schemas | JSON | JSON | Orchestration | Python loops | Custom graphs | Workflow graphs | Sequential/parallel | Memory | SQLite typed, relational | External DB | File-based | Vector DB | Sandboxing | OpenShell integration | Manual | Manual | Manual | Performance | ✅ 2x token efficiency | ❌ Higher overhead | ❌ Moderate | ❌ Moderate | Inspectability | ✅ Single class | ❌ Scattered configs | ❌ Mixed abstractions | ❌ JSON-heavy | Core framework pip install nooa With memory and CLI tools pip install "nooa memory,cli " nooa --version Should output = 0.1.0 scanner.py . docker run -it --rm -v $ pwd :/app nvcr.io/nvidia/openshell:latest python /app/scanner.py nooa trace . sqlite3 agent memory.db "SELECT FROM memories LIMIT 5;" mypy , pytest . “NOOA proves that the harness around a model matters as much as the model itself.”—NVIDIA Research pdb or test with pytest .NOOA is a paradigm shift in AI agent development: For developers building cybersecurity tools, DevOps assistants, or research agents , NOOA offers a rare blend of power and simplicity . As the framework matures, expect it to influence how we test, deploy, and trust AI systems. Have you built a NOOA agent? Share your use case in the comments Cover image suggestion: A side-by-side comparison of NOOA’s Python class vs. traditional JSON-based agent configurations, or a diagram of the VulnScannerAgent workflow. ~TrixSec