Agent Skills Quickstart

NVFLARE Agent Skills provide guided workflows for understanding NVFLARE, creating jobs, optimizing an existing job, and reporting AutoFL results. You describe the goal in natural language; the coding agent selects the appropriate installed skill from the request and available source files.

Install the Skills

Install all skills from an NVFLARE checkout:

npx skills add ./skills --skill '*' -a codex -a claude-code -y

The generated jobs also require NVFLARE 2.9.0 or later in the Python environment used by the coding agent. Install the package from PyPI or use an editable NVFLARE checkout.

Choose a Workflow

You do not need to know or name the internal skill. Choose the goal that matches your task and provide the relevant project, data, job, or evidence.

Understand Where to Start

Use the orientation workflow when you are new to NVFLARE, are unsure which API or workflow fits a project, or have source code with more than one training framework or entry point. It inspects the supplied project read-only and recommends one next workflow without editing files or starting a job.

Create a Federated Job

Use the job-creation workflows for either of these goals:

  • Convert existing plain PyTorch, PyTorch Lightning, or Hugging Face Trainer code into a federated learning training job.

  • Given a sample tabular or image dataset, generate a federated statistics job.

See Agent Conversion Skills for supported inputs, example prompts, generated results, validation behavior, and limitations.

Optimize an Existing Job

Use AutoFL when you already have a runnable NVFLARE job.py and want to improve a measured objective through reproducible candidate experiments. State the metric, execution environment (simulation, POC, or production), and any candidate limit. AutoFL optimizes an existing job; it does not convert standalone training code.

See NVFlare Auto-FL Agent Skill for campaign behavior, permissions, comparison rules, and outputs.

Report a Completed AutoFL Campaign

Use AutoFL reporting after a campaign has stopped, reached its candidate cap, or been explicitly interrupted. It turns the recorded campaign evidence into a reproducible Markdown report, JSON summary, and progress plot without changing campaign results.

Try the Runnable Examples

The Agent Skills runnable examples provide five small, standalone projects for trying NVFLARE Agent Skills with Codex or Claude Code. Each includes starting source or synthetic data, an exact natural-language prompt, and a local validation path.

Open one example directory in your coding agent and paste its prompt. The examples cover plain PyTorch, PyTorch Lightning, Hugging Face Trainer, tabular federated statistics, and image federated statistics. Review the proposed changes before running a simulation or applying a workflow to real data.