Wandb Skills
Official Weights and Biases skills: an agent analyzes experiments and traces, instruments Weave and runs training research.
- 75 ★ on the repository
- Last commit
- Apache-2.0
Extended access
Why this level
- The agent receives a Weights and Biases key with access to projects and runs.
- The `wandb-autoresearch` skill launches training through Launch.
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add skills/wandb-skillsDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The agent receives a Weights and Biases key with access to projects and runs.
- The `wandb-autoresearch` skill launches training through Launch.
This is third-party code. Review the repository files before installing.
What it does
Official Weights and Biases skills for agents that train models and build AI applications. Five skills: `wandb-primary` analyzes projects, runs, artifacts and reports, `wandb-eval-tables` turns tables into verified evaluation previews, `wandb-autoresearch` runs bounded training research through Launch, and `weave-instrument` and `agentlens-tracing-instrument` add tracing to Python and TypeScript apps. The agent learns to log metrics and media, compare experiments, tune hyperparameters, group traces by failure mode and run online evaluations. The README reports the project's own Skill Bench results for Claude Code and Codex.
Who it is for. Machine learning engineers and agent developers with a Weights and Biases account who hand run and trace analysis to Claude Code or Codex.
Good fit when
- A loss curve looks odd and the agent compares runs of the project.
- An LLM application runs without tracing and Weave instrumentation is added.
- Traces are grouped by failure cause before editing a prompt.
- Hyperparameter tuning runs as a bounded series through Launch.
- Evaluation results sit in a Weights and Biases table and need a verified preview.
Not a fit when
- Experiments live in another system without Weights and Biases.
- The Weights and Biases key cannot be given to an agent.
- The machine has Python older than 3.13 and cannot be upgraded.
Example request
Compare the last three runs of the project, find why validation loss diverged and suggest which hyperparameters to change.Limitations
A Weights and Biases account and a key in `WANDB_API_KEY` are required. The helper scripts are validated on Python 3.13, `wandb` 0.28.1 and `weave` 0.52.41; older versions leave some methods unimplemented. Workspace helpers need the extra `wandb[workspaces]` package. All skills in the README table are marked experimental. Installation adds instructions only, and you supply dependencies and keys yourself. README benchmark results for Codex are lower than for Claude Code.
How to disable. Delete the `wandb-*`, `weave-instrument` and `agentlens-tracing-instrument` folders from the agent's skills directory. Remove `WANDB_API_KEY` from the environment.
FAQ
Which skills are included?
`wandb-primary`, `wandb-eval-tables`, `wandb-autoresearch`, `weave-instrument` and `agentlens-tracing-instrument`. Pick one with `--skill <name>`.
How do I get a key?
Create it at `https://wandb.ai/authorize` and pass it with `export WANDB_API_KEY=<key>`.
Which versions are needed?
Python 3.13 or newer, `wandb` 0.28.1 or newer and `weave` 0.52.41 or newer. Version 0.28.0 is not enough: some helpers rely on later changes.
How do I run a script with extra packages?
Through `uv run --with 'wandb[workspaces]>=0.28.1' --with 'weave>=0.52.41' python your_script.py`.
Does the set suit every agent?
The `npx skills` installer works with major coding agents; `--global` and `--agent <name>` choose scope and target agent.
Discussion
No questions or reviews yet.
Related
SkillsA CLI that runs your agent on tasks and revises its skills, keeping a change only when held-out tasks confirm the gain
A catalog of nearly two hundred skills: code delivery, review, debugging, AI governance, product work and agent coordination.
A collection of 14 portable skills for long agent runs: work planning, live app QA, search across past sessions and a phone call
A skill for the "Early Experience" method (ICML 2026) that helps an agent prepare reward-free training data for language agents.
A skill that designs, tests, evaluates and packages other agent skills, starting from architecture
90 specialist skills and the Nexus orchestrator for Claude Code, Codex and Antigravity: bugs, tests, security, review, UX
README in short
The set divides into model training and agent building. Tasks are listed per group, followed by version requirements and a skill table with status.