Start Here: Run Your First Agent
Run your first code review with NVIDIA NeMo Fabric. Give an agent a small Python function, inspect the bug it finds, and read the result returned through NeMo Fabric. Choose the starter notebook for an interactive walkthrough or the Python Quickstart for a script you can extend.
Use the Starter Notebook
Prepare Your Checkout
The starter notebook uses Hermes Agent from a NeMo Fabric source checkout. Complete Development Setup and the supported Hermes Agent installation, then provide an NVIDIA API key. The notebook setup guide explains the notebook’s environment and credential prompts.
Open or download the starter notebook on GitHub. A source checkout includes it at examples/notebooks/01_quickstart.ipynb.
Open the Notebook in Jupyter
From the prepared repository root, run:
Open 01_quickstart.ipynb and run its cells in order. The notebook constructs the reviewer and supplies this function directly in the request; it does not need to read project files:
Recognize a Successful Review
With the notebook’s prerequisites available, the run should return status succeeded. The review should identify that an empty input list causes division by zero. Wording and suggested fixes can vary; check both the status and the finding. A skipped live cell means a prerequisite is missing, not that a review completed.
The remaining cells demonstrate runtime reuse and optional diagnostics. Notebook 02_variations.ipynb explores harness and configuration changes after the first run.
Follow the Python Quickstart
These three connected guides use the same code-review task. Follow them in order; the later guides reuse the configuration from the first. An optional fourth guide moves to the source-checkout OpenShell example:
- Install One Harness. Install NeMo Fabric with LangChain Deep Agents, provide model access, and review the function. Keep the resulting
first_agent.pyfor the next two guides. - Switch Harnesses. Reuse that configuration to run Deep Agents and Hermes Agent variants, then compare results through the same application interface.
- Add Skills and MCP. Extend the Hermes variant with a review skill and a working local MCP tool, then inspect evidence of its tool calls.
- Run in OpenShell (Optional). On a Linux host, run the source-checkout Hermes code-review example in a sandbox.
Start with Python 3.11 or later and an NVIDIA API key. The second guide adds a separate Python 3.14 environment for Hermes Agent; the third supplies a complete local stdio MCP server. The optional sandbox guide requires a Linux host and a source checkout. Each guide explains its additional setup. For supported platforms and other installation routes, see the Installation Overview.