NeMo Agent Toolkit
How to migrate from the NVIDIA NeMo Agent Toolkit
Move one capability at a time. List which parts of the NeMo Agent Toolkit you use, then move tracing to NeMo Relay, evaluation to NeMo Helix and agent execution to NeMo Fabric, in that order. Run the same evaluation set before and after each step so you can show nothing got worse. NVIDIA recommends completing the move before the end of 2026.
Before you start
- Pin the toolkit version you run today, so nothing changes under you during the move.
- Write down which toolkit features each agent uses.
- Make sure you have an evaluation set for each agent. It is your proof that the migrated agent behaves the same.
- Read the migration guide NVIDIA publishes with the toolkit documentation.
Map what you use
| Toolkit feature | Moves to | Notes |
|---|---|---|
| Tracing and telemetry export | NeMo Relay | Exports to OpenTelemetry and OpenInference |
| Middleware and policy on calls | NeMo Relay | Middleware can block, redact or rewrite a call |
| Token, latency and cost measurement | NeMo Relay | Through its plugins |
| Evaluation runs | NeMo Helix | Benchmarks, test sets and scoring |
| Prompt and configuration optimization | NeMo Helix | |
| Agent configuration and execution | NeMo Fabric | Check the adapter compatibility matrix first |
| Agent logic written in a framework | Stays where it is | LangGraph, CrewAI, LangChain and Google ADK code is yours |
A safe order
- Step 1, tracing. Add NeMo Relay beside the toolkit and send traces to the same observability tool. The toolkit includes an experimental bridge to Relay telemetry, which makes this the lowest-risk first step.
- Step 2, evaluation. Recreate your evaluation set in NeMo Helix and confirm the scores match what the toolkit reported.
- Step 3, execution. Move one agent to NeMo Fabric, choosing an adapter that supports the tools and settings it needs.
- Step 4, compare. Run the evaluation set against the old and new versions of that agent.
- Step 5, repeat for the remaining agents, then remove the toolkit.
Risks to plan for
- The successors were pre-1.0 releases in November 2026. Interfaces may change between releases, so pin versions.
- Support differs by adapter in NeMo Fabric. A feature your agent relies on may not be available in every adapter.
- NeMo Relay is a runtime with software kits for Python, Node.js and Rust, not a drop-in Python package swap.
- Disconnected environments need the new packages and images mirrored inside before the move.
- Evaluation scores can shift when the evaluator changes. Compare like with like.
If you built on a platform
The work above falls on whoever owns the integration with the library. If your agents were built on a platform, that is the platform vendor. Ask the vendor three things: which successor projects they are moving to, when, and whether your agents need any change.
Where Katonic fits
Every agent built in Katonic is a declarative definition that compiles to NVIDIA's open agent software. Today that software is the NeMo Agent Toolkit. When NVIDIA changes the runtime, Katonic changes the compiler and your agents carry over. The definition of the agent, its knowledge, tools, guardrails and approval points, does not change.
Read next
- NVIDIA NeMo Agent Toolkit is deprecated: what replaces it? →
- NeMo Relay vs NeMo Fabric vs NeMo Helix →
- How to evaluate agents built with the NeMo Agent Toolkit →
- NVIDIA migration guide →
Questions and answers
When does the NeMo Agent Toolkit reach end of life?
By the end of 2026. NVIDIA deprecated it as of October 2026 and recommends completing migration before the end-of-life date.
Do I have to rewrite my agents?
Not the agent logic written in a framework such as LangGraph or CrewAI. What changes is the software around it: tracing, evaluation and how the agent is run.
What should I migrate first?
Tracing, to NeMo Relay. It can run beside the toolkit, so it carries the least risk.
Can I keep using the toolkit after end of life?
It is open source and will keep running, but you should not expect fixes, including security fixes.
How do I know the migrated agent behaves the same?
Run the same evaluation set against the old and new versions and compare the scores per item.
Do agents built in Katonic need to be migrated?
No. They are declarative definitions, and Katonic changes the compiler when NVIDIA changes the runtime.