arbor

SUMMARY

What it does

This skill enables autonomous optimization of a concrete artifact (code, training recipe, agent harness, data pipeline, prompt) against a measurable objective and evaluator, using Hypothesis Tree Refinement (HTR) from the Arbor paper. It is designed for iterative improvement tasks involving many experiment-evaluate cycles, such as raising model eval scores, improving agent harnesses, tuning pipelines, beating baselines on benchmarks, or MLE-bench/Kaggle-style optimization. The skill orchestrates Claude as a coordinator with subagent executors in isolated git worktrees, maintaining a persistent hypothesis tree to accumulate insights and avoid overfitting. It requires a dev/test evaluator split and uses a held-out merge gate to admit changes only when they improve on the test evaluator. The skill includes scripts for tree management and references for methodology and executor briefs. It is intended for long-horizon tasks where the bottleneck is organizing many trials rather than writing a single change.