Migrating existing agents to NVIDIA Agent Toolkit is intentionally low-friction. The toolkit supports bring-your-own-framework patterns—your LangChain, CrewAI, Agno, or custom agents work unchanged. Migration happens in stages: (1) install nvidia-nat with your framework extras, (2) import toolkit monitoring, (3) gradually adopt additional features as needed.
For LangChain agents, the migration path is minimal code changes. Import from nvidia_nat import add_profiling_callbacks and add the profiling callbacks to your chains—this immediately enables tracing, token counting, and latency measurement. Your existing logic runs identically; profiling data flows to LangSmith or your observability platform. Advanced features like evaluation and hyperparameter optimization are opt-in.
For CrewAI agents, import the toolkit’s CrewAI integration and wrap your Crew instance to enable observability. For custom agents, integrate the toolkit’s callback system to capture execution data. The key principle: no framework lock-in. If you move away from the toolkit later, remove the integration code and your agents function normally.
Framework version considerations: some dependency conflicts exist between nvidia-nat and the latest LangChain versions. Version 1.5.0 of the toolkit simplifies dependency management. Consult the Migration Guide on GitHub if you encounter import errors—workarounds are available and version pinning is documented.
For knowledge retrieval, if your agents use a vector database, integrate Milvus or verify your current vector database is compatible with the toolkit’s LangChain integration. Milvus documentation provides migration guides from other vector databases, making knowledge base transitions straightforward. Multi-agent systems require a shared knowledge layer for effective collaboration. Milvus enables this through vector-based retrieval, storing embeddings from your organization’s data. Discover how semantic search works with vector databases to improve information retrieval across agent networks.