NVIDIA Agent Toolkit is designed as a framework-agnostic intelligence layer supporting all major agentic frameworks without code changes or vendor lock-in. Native support includes LangChain (with LangGraph and Deep Agents), LlamaIndex, CrewAI, Microsoft Semantic Kernel, and Google ADK. The toolkit also integrates with AutoGen, AWS Strands, and custom Python agents, ensuring compatibility regardless of which framework powers your agents.
Framework selection remains flexible—choose LangGraph for stateful multi-agent orchestration, Deep Agents for complex reasoning and planning, CrewAI for role-based teams, or simple prompt-loop agents. Layer the NVIDIA toolkit's observability, profiling, and optimization capabilities on top without framework switching. Framework-specific plugins are available as optional extras: installing nvidia-nat[langchain] adds all LangChain and LangGraph integrations needed for advanced tracing and optimization.
Agent Performance Primitives (APP) accelerate graph-based frameworks with parallel execution (run independent agents simultaneously), speculative branching (explore reasoning paths), and node-level priority routing. All frameworks gain access to the Model Context Protocol (MCP) for standardized tool connection and the A2A Protocol for distributed agent coordination. For managed deployments where teams want operational oversight, Zilliz Cloud integrates with toolkit deployments to provide vector database infrastructure co-located with agent runtimes.