NVIDIA shows how to train a safe CLI agent with synthetic data + RLVR
NVIDIA published a tutorial today on training a command-line agent for the LangGraph Platform CLI using synthetic data generation and reinforcement learning with verifiable rewards. The walkthrough centers on the Nemotron-Nano-9B-V2 base model and a confirm-before-execute interface.
NVIDIA published a tutorial today on training an AI agent to operate the LangGraph Platform CLI using synthetic data generation (SDG) and reinforcement learning. The approach trains the model to propose a CLI command, request a human yes/no confirmation, and then report results such as a server starting on port 8000.
In the data step, NVIDIA uses NeMo Data Designer to expand a small set of seed examples into hundreds of natural-language requests paired with LangGraph CLI invocations, then exports the dataset in an OpenAI-style messages format. For training, the post describes Reinforcement Learning with Verifiable Rewards (RLVR) where a deterministic verifier enforces constraints like starting commands with langgraph and limiting subcommands to options including dev, up, build, and dockerfile. The prerequisites list Python 3.10+, CUDA 12.0+, at least 32 GB system RAM, about 100 GB of free disk, and an NVIDIA GPU with 80 GB memory (for example, an A100).