# NVIDIA NeMo Agent Toolkit NVIDIA NeMo Agent Toolkit (`nvidia-nat`) is an open-source [[Python]] library for connecting, evaluating, and accelerating teams of [[AI Agents]]. Part of the [[NVIDIA Agent Toolkit]]. It is a middleware layer that adds intelligence across any agent framework rather than replacing existing ones. Apache 2.0 license, Python 3.11+. ## Four pillars 1. **BUILD**: YAML-configured workflows with modular functions, LLM providers (NIM, OpenAI, Azure, HuggingFace, Bedrock, LiteLLM), retrievers for [[Retrieval-Augmented Generation (RAG)]], memory providers, MCP client, and Agent-to-Agent (A2A) protocol support 2. **RUN**: CLI, API server + web UI, MCP server, A2A server, [[LangGraph]] integration. Observability via Phoenix, Langfuse, LangSmith, OpenTelemetry 3. **IMPROVE**: evaluation system, profiling (agent to token level), hyperparameter and prompt optimization, RL fine-tuning via OpenPipe ART and DPO with NeMo Customizer 4. **EXTEND**: plugin system for custom functions, agent types, auth providers, LLM providers, retrievers, evaluators, memory providers ## Agent types Eight built-in: ReAct, Reasoning, ReWOO (plans all tool calls upfront), Responses API, Router (task dispatcher), Sequential Executor, Tool Calling, and Automatic Memory Wrapper. ## Framework integrations Works alongside [[LangChain]]/[[LangGraph]], LlamaIndex, CrewAI, Microsoft Semantic Kernel, Google ADK, Microsoft AutoGen. Each integration provides callback handlers, LLM adapters, and tool wrappers. ## Security Defense middleware for content guards, PII detection, output verification. Red teaming middleware for vulnerability identification. Detects prompt injection, jailbreak attempts, and tool poisoning. ## References - https://docs.nvidia.com/nemo/agent-toolkit/latest/ - https://github.com/NVIDIA/NeMo-Agent-Toolkit ## Related - [[NVIDIA Agent Toolkit]] - [[AI Agents]] - [[AI Agent Orchestration]] - [[NeMo]] - [[NVIDIA Nemotron]] - [[Retrieval-Augmented Generation (RAG)]] - [[LangChain]] - [[LangGraph]] - [[Python]]