NVIDIA NeMo Agent Toolkit Release Notes#

This section contains the release notes for NeMo Agent toolkit.

Release 1.4.0#

Summary#

This release introduces initial support for several frameworks and integrations including A2A, AWS Strands, Amazon Bedrock AgentCore, Microsoft AutoGen, and NVIDIA Dynamo. In addition to new framework and integrations, an automatic agent wrapper for LangGraph Agents enables users to bring their own agent. Per-user functions enable deferred instantiation which provides per-user stateful functions, per-user resources, and other useful features. The toolkit continues to offer backwards compatibility, making the transition seamless for existing users.

  • LangGraph Agent Automatic Wrapper: Easily onboard existing LangGraph agents to NeMo Agent Toolkit. Use the automatic wrapper to access NeMo Agent Toolkit advanced features with very little modification of LangGraph agents.

  • Automatic Reinforcement Learning (RL): Improve your agent quality by fine-tuning open LLMs to better understand your agent’s workflows, tools, and prompts. Perform GRPO with OpenPipe ART or DPO with NeMo Customizer using NeMo Agent Toolkit built-in evaluation system as a verifier.

  • Initial NVIDIA Dynamo Integration: Accelerate end-to-end deployment of agentic workflows with initial Dynamo support. Utilize the new agent-aware router to improve worker latency by predicting future agent behavior.

  • A2A Support: Build teams of distributed agents using the A2A protocol.

  • Safety and Security Engine: Strengthen safety and security workflows by simulating scenario-based attacks, profiling risk, running guardrail-ready evaluations, and applying defenses with red teaming. Validate defenses, profile risk, monitor behavior, and harden agents across any framework.

  • Amazon Bedrock AgentCore and Strands Agents Support: Build agents using Strands Agents framework and deploy them securely on Amazon Bedrock AgentCore runtime.

  • Microsoft AutoGen Support: Build agents using the Microsoft AutoGen framework.

  • Per-User Functions: Use per-user functions for deferred instantiation, enabling per-user stateful functions, per-user resources, and other features.

Refer to the changelog for a complete list of changes.

Known Issues#