skills

Skills to guide Claude Code, Codex, and other coding agents on using the Weights & Biases AI developer platform to train models and build agents.

For model training

• Log metrics and rich media during model training and fine-tuning
• Track model training experiments
• Analyze runs and experiment results to understand how the model is learning
• Tune hyperparameters

For agent building

• Trace agentic AI applications
• Analyze traces and classify them into failure modes
• Evaluate models with labeled datasets
• Run online evaluations for production monitoring

Getting Started

npx skills add wandb/skills

Then set your W&B API key:

export WANDB_API_KEY=<your-key>

npx skills is a utility for installing skills into major coding agent CLIs. Use --global to install for all projects, or --agent <name> to target a specific agent. See the npx skills docs for more details.

Requirements

The skill helpers are validated against these versions. Older SDKs expose some of the methods used here as unimplemented stubs, so pin at or above the floor rather than relying on whatever is already installed.

Package	Floor	Validated against
Python	>=3.13	3.13
wandb	>=0.28.1	0.28.1
wandb-workspaces	>=0.4.4	0.4.4 (pulled in by the [workspaces] extra)
weave	>=0.52.41	0.52.41

The workspaces extra is required for the Workspaces helpers:

uv run --with 'wandb[workspaces]>=0.28.1' --with 'weave>=0.52.41' python your_script.py

0.28.1 is a floor rather than a preference: at least one helper branches on behavior that changed after 0.28.0, so >=0.28.0 is not sufficient.

Available Skills

Skill	Description	Status
wandb-primary	Broad W&B project analysis and operations across runs, Artifacts, Registry, Weave, Reports, Workspaces, and Launch.	experimental
wandb-eval-tables	Non-destructive conversion of W&B Table artifacts into bounded, verified EvalTable previews.	experimental
wandb-autoresearch	Bounded training research through W&B Launch, including readiness checks, serial trials, comparison, and resumable state.	experimental

Benchmarks

We maintain Skill Bench in this repository to evaluate public skill changes across coding agents and task categories. Skill Bench uses W&B Agent Factory as the eval runtime for task definitions, agent profiles, sandbox execution, and structured bench rows.

Pull requests run package validation by default. A maintainer can trigger live Skill Bench runs for larger changes.

Plan a local benchmark without model calls:

python3 -m skillbench.cli plan \
  --wbaf-root ../WandBAgentFactory \
  --candidate-ref HEAD \
  --skill wandb-primary

Category	Tasks	Claude Code (sonnet4.6)	Codex (gpt-5.3-codex)
Weave analysis	26	97%*	63%*
Weave tooling	11	95%*	83%*
Model training	8	90%*	85%*
LLM finetuning & RL analysis	14	72%*	86%*
Failure & outlier detection	8	86%*	63%*

*Pass rates are +/- 3%. Many tasks span multiple categories.

Contributing

See CONTRIBUTING.md.