Skip to content
Get packs

Guide · Agent skill

How to use the Weights & Biases experiment tracking agent skill

Updated · written with AI from the project's own documentation and checked against our catalogue data

This guide explains how to install and use the weights-and-biases agent skill, which teaches your AI agent how to add Weights & Biases (W&B) experiment tracking to machine learning code. It is written for people who have not used agent skills before. By the end you will know what the skill changes, how to install it, how to ask for the work in plain language, and what you need in place (Python, the wandb package and a W&B account with an API key).

Weights & Biases Experiment TrackingOrchestra-Research/AI-Research-SKILLs · MIT

Quick start

  1. 1Download and unzip the skill. Download the ZIP from this page and unzip it into ~/.claude/skills/ or your project's .claude/skills/ folder. Or upload it under Settings → Capabilities → Skills in the Claude apps.
  2. 2Install wandb. Run pip install wandb in the Python environment your training code uses.
  3. 3Connect your account. Run wandb login, or set WANDB_API_KEY to your API key.
  4. 4Ask for the tracking work. Describe your script and goal, for example adding W&B logging to train.py.
  5. 5Review the generated code. Read the wandb.init, wandb.log and artifact calls before running them.
  6. 6Run and check the dashboard. Run your script and open the run in the W&B dashboard to see metrics and compare runs.

What does the weights-and-biases skill make the agent do?

With the skill installed, the agent knows how to add Weights & Biases, a hosted tool for machine learning experiment tracking, to a training script. It can write code that starts a run, records settings (config) and metrics such as loss and accuracy, saves checkpoints and datasets as versioned artifacts, and sets up automated hyperparameter sweeps.

The results appear in an online dashboard where runs can be compared and shared with a team.

Install the skill (ZIP download or upload)

Download the ZIP file for the skill from this page and unzip it. For a personal setup, place the unzipped folder in ~/.claude/skills/. For a single project, place it in that project's .claude/skills/ folder.

You can also upload the skill in the Claude apps under Settings → Capabilities → Skills.

  • The skill folder contains SKILL.md, plus references/artifacts.md, references/integrations.md and references/sweeps.md.
  • Install the wandb Python package before running the code the agent writes.
pip install wandb

Set up W&B login and your API key

The skill needs a W&B account and API key. You can run wandb login to create and store a key, or set the WANDB_API_KEY environment variable yourself.

# Login (creates API key) wandb login # Or set API key programmatically export WANDB_API_KEY=your_api_key_here

Trigger the skill with a natural request

You do not need special syntax. Describe the task in your own words, and name the thing you want, such as experiment tracking, a sweep or an artifact. The agent should pick up the skill when your request matches its description.

Good requests name your script and the outcome. For example: add W&B logging to train.py so loss and validation accuracy appear on a dashboard, or set up a Bayesian sweep over learning rate and batch size that maximizes val/accuracy.

  • Add wandb.init and wandb.log to my PyTorch training loop in train.py.
  • Save my model checkpoint as a W&B artifact after each best validation run.
  • Create a grid sweep over learning rate and batch size and run 20 trials.

What the agent's code typically looks like

A basic run starts with wandb.init, passes a config dictionary, logs metrics inside the training loop and ends with wandb.finish. Sweeps use a sweep_config with a method (grid, random or bayes), a metric and parameters, then call wandb.agent to run the trials.

Artifacts are created with wandb.Artifact, given files with add_file or add_dir, and sent with wandb.log_artifact. Teammates or later runs can fetch them with run.use_artifact.

  • Grid: exhaustive search over the listed values.
  • Random: samples from the distributions you define.
  • Bayesian (bayes): the skill recommends it for most sweeps.
wandb.sweep(sweep_config, project="my-project") wandb.agent(sweep_id, function=train, count=50)

Safety notes: scripts, network and privacy

This skill runs scripts and uses the network. The workflow sends metrics and uploaded files to W&B servers and requires an account and API key. Do not use it where training logs must stay fully offline or private.

Check any code the agent writes before running it, especially where it uploads files, and keep your API key out of shared code.

  • Requirements: Python, the wandb package, a W&B account and API key, and network access to W&B servers.
  • Not suitable for non-ML software work; the skill only covers machine learning experiment tracking.

What people use it for

  • Logging loss and accuracy from a PyTorch training loop to a shared dashboard.
  • Running a Bayesian hyperparameter sweep that maximizes validation accuracy.
  • Versioning a training dataset or model checkpoint as a W&B artifact so teammates can download the same files.
  • Comparing runs with different configs, tags and notes in one project.

Frequently asked questions

How do I install the weights-and-biases skill?

Download the skill ZIP from this page and unzip it into ~/.claude/skills/ for personal use, or into a project's .claude/skills/ folder. You can also upload it in the Claude apps under Settings → Capabilities → Skills.

What do I need to use this skill?

You need Python, the wandb package (pip install wandb), a W&B account with an API key set through wandb login or WANDB_API_KEY, and network access to W&B servers.

Is Weights & Biases free?

The source does not state pricing or licence terms for the W&B service. The skill itself is listed under the MIT licence. Check W&B's own site for account and plan details.

Can I use this skill without sending data to the internet?

No. The workflow sends metrics and uploaded files to W&B servers and requires an account and API key, so it is not suitable for fully offline or private logging.

What are the alternatives to this skill?

The source does not compare this skill with other tools. You can still ask your agent to write experiment-tracking code without it, but this skill specifically covers W&B.

Why isn't my agent using the skill?

Check that the folder is in ~/.claude/skills/ or the project's .claude/skills/ and that SKILL.md is inside it. Then phrase your request around experiment tracking, sweeps or artifacts so it matches the skill's description.

New agent skills, repos and AI tools — every week

The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.

Our verdict

Use this skill if you are an ML engineer or researcher who wants an agent to add W&B run tracking, artifacts and hyperparameter sweeps to your code and review results in a web dashboard. Pick something else if your training logs must stay offline or private, or if your work is not machine learning.

Download the skill