ML engineer
Agent Plugins 1.0.0 compatible
Take models from experiment to production with offline evals, experiment tracking, training run debugging, serving readiness, and model cards.
Version 1.0.2Updated Oct 20, 2026Source
First time? Install with npm i -g harnesstap
Download as Agent Plugins zipWhat's inside
Skills (6)
dataset-versioningVersion training and evaluation datasets with immutable snapshots, content hashes, and documented lineage so every model can be traced to the exact data it saw. Use when creating a new dataset, refreshing training data, fixing labels, or when an eval set may have changed silently.eval-harnessBuild an offline evaluation harness that scores models on fixed, versioned datasets with task-appropriate metrics, slices, and baselines, so model changes are compared fairly. Use when training or fine-tuning a model, comparing candidates, or gating promotion. Agent trajectory evals belong to agentic-ai-engineer.experiment-trackingSet up and use experiment tracking so every training run records code version, data version, config, metrics, and artifacts, and runs can be compared and reproduced. Use when starting a modeling project, when results cannot be reproduced, or when nobody knows which run produced the production model.model-cardWrite a model card documenting what a model does, its training data, evaluation results by slice, limitations, risks, and intended and out-of-scope uses. Use before promoting a model to production, publishing it, or handing it to another team.serving-readinessCheck that a model is ready to serve in production: latency and throughput under load, cost per prediction, input validation, versioning, fallback behavior, monitoring, and rollback. Use before first deployment of a model or a major model change.training-run-debuggingDiagnose training runs that diverge, plateau, overfit, run out of memory, or produce suspiciously good results, by checking data, configuration, and optimization in a fixed order. Use when loss curves look wrong, metrics do not improve, or a run crashes.
Rules (3)
gpu-budgetUse GPU time deliberately and within budget.inference-sloEvery production model has latency, availability, and quality targets with alerts.training-reproducibilityEvery training run must be reproducible from its record.
Commands (2)
promote-modelCheck promotion gates for a model version and prepare its promotion in the registry.run-evalRun the offline eval harness for a model on a dataset version and compare with production.
Plugins (5)
agentic-bundle-aas-data-engineering-platform@agentic-awesome-skillsfeature-dev@claude-plugins-officialcontext7@claude-plugins-officialagent-skills@addy-agent-skillssuperpowers@claude-plugins-official