Object-Oriented Agents

Summary

Object-Oriented Agents (NOOA) is an open-source, Apache 2.0-licensed Python framework for building AI agents, distributed as a library you install with `pip install nooa` or `uv add nooa` and import into your own code. Rather than treating prompts, tools, callbacks, and workflows as separate abstractions, as many agent frameworks do, it lets a developer express an agent's state, capabilities, prompts, and typed interfaces as a single Python class, which makes it a fit for application developers building agentic systems rather than for security teams directly, though an included cybergym example shows it used to build a cybersecurity-oriented agent. Recent hardening work has removed an eval()-based code execution path in its CodeAct constructor, confined shell-tool file operations to the working directory, and added auth and CORS restrictions to its local trace viewer, work that suggests active, ongoing maintenance.

NVIDIA Object Oriented Agents: the Pythonic way to build AI Agents.

Command line

Values are placeholders taken from each option’s declared default. Nothing is executed here — the output shown is a recording of a run that already happened.

Release history

1. v0.0.9 Aug 18, 2026 · issue -001

   OO-Agents v0.0.9 adds plugin-based CLI extension, a portable journal exporter, callable LLM routing in @strategy, shared interactive sessions, and a CyberGym portfolio agent.

   └──▷ GET THIS VERSION
   copy

   $ git clone --branch v0.0.9 https://github.com/NVIDIA-NeMo/labs-OO-Agents.git
   # already have the repo? check out this version:
   $ git checkout v0.0.9

   • ›Enables external packages to add nooa subcommands via Python entry points, allowing third-party plugins to extend the CLI without forking.
   • ›Adds a portable journal file exporter for tracing, giving teams an offline-portable trace format independent of a remote viewer.
   • ›Allows a callable to be passed to @strategy(llm=...), enabling dynamic LLM selection per strategy invocation instead of a static reference.
   • ›Adds LLM routing metadata to generation observability, surfacing which backend was selected in trace output.
   • ›Adds a shared interactive session foundation, enabling multiple agents to collaborate within a single interactive context.
   +1 moreshow less

   • ›Adds a portfolio-based NOOA agent to the CyberGym benchmark environment.
   └──▷ BREAKING ON UPGRADE

   • !The nemo_flow_* namespace is renamed to nemo_relay_*; any code or config referencing nemo_flow_ names will break on upgrade.
2. v0.0.9 Aug 18, 2026 · issue 001

   OO Agents v0.0.9 adds entry-point CLI extensibility, a callable @strategy(llm=...), a portable journal exporter, shared interactive sessions, and a CyberGym portfolio agent.

   └──▷ GET THIS VERSION
   copy

   $ git clone --branch v0.0.9 https://github.com/NVIDIA-NeMo/labs-OO-Agents.git
   # already have the repo? check out this version:
   $ git checkout v0.0.9

   └──▷ USE IT
   Dynamically select an LLM per-call (e.g. based on task complexity) instead of binding one model at class definition time.
   python copy

   from nooa import Agent
   from nooa.unifiedllm import get_llm_client
   
   def pick_llm():
       return get_llm_client("ollama_chat/qwen3:1.7b", api_base="http://localhost:11434")
   
   class AdaptiveAgent(Agent):
       @strategy(llm=pick_llm)
       async def analyze(self, text: str) -> str:
           """Analyze the text."""
           ...

   • ›Allows a callable to be passed to @strategy(llm=...) so LLM selection can be decided at runtime rather than at class-definition time.
   • ›Adds a portable journal file exporter for tracing, enabling offline and cross-environment trace storage.
   • ›Lets external packages register new nooa subcommands via Python entry points, making the nooa CLI extensible without modifying core.
   • ›Adds a portfolio-based NOOA agent to the CyberGym benchmark environment.
   • ›Adds LLM routing metadata to generation observability, surfacing which backend handled each LLM call in traces.
   └──▷ BREAKING ON UPGRADE

   • !All nemo_flow_* symbols are renamed to nemo_relay_*; any code importing or referencing nemo_flow_* names will break.
3. v0.0.7 Jul 30, 2026 · issue -020

   v0.0.7 adds a CyberGym agent example, viewer ingest resource limits, and playground custom-model endpoint constraints.

   └──▷ GET THIS VERSION
   copy

   $ git clone --branch v0.0.7 https://github.com/NVIDIA-NeMo/labs-OO-Agents.git
   # already have the repo? check out this version:
   $ git checkout v0.0.7

   • ›Constrains playground custom-model endpoint and api_key_env to server-declared pairs, preventing arbitrary endpoint injection.
   • ›Adds resource limits on the Viewer ingest API to cap payload sizes and prevent abuse.
   • ›Adds a nooa CyberGym agent example demonstrating object-oriented agent patterns for cyber exercise environments.
   • ›Requires an auth token and restricts CORS on the viewer API, hardening the exposed surface.
   • ›Constrains ShellTools file operations to the current working directory, limiting lateral file access.
   +1 moreshow less

   • ›Replaces eval() in the CodeAct constructor-string coercion path with an AST decoder, removing arbitrary code execution risk in that call site.
4. v0.0.7 Jul 30, 2026 · issue 001

   NOOA v0.0.7 adds auth-gated viewer API, secret redaction in telemetry, resource limits on viewer ingest, and a NOOA CyberGym agent example.

   └──▷ GET THIS VERSION
   copy

   $ git clone --branch v0.0.7 https://github.com/NVIDIA-NeMo/labs-OO-Agents.git
   # already have the repo? check out this version:
   $ git checkout v0.0.7

   • ›Adds authentication token requirement and CORS restrictions to the viewer API, preventing unauthenticated access to trace data.
   • ›Constrains playground custom-model endpoint and api_key_env fields to server-declared pairs, blocking arbitrary endpoint injection.
   • ›Enforces resource limits on viewer ingest to bound memory and CPU consumption under high trace volume.
   • ›Confines ShellTools file operations within the current working directory, reducing the blast radius of LLM-generated shell actions.
   • ›Adds an NOOA CyberGym agent example demonstrating how to build a cyber-domain agent with the OO Agents framework.