What problem does it solve? Working on the pharma-agents project requires understanding a complex autonomous ML optimization loop where CrewAI agents propose, implement, and evaluate changes to molecular property prediction models. This Skill provides the architectural knowledge, file conventions, and debugging patterns needed to run experiments, interpret results, and fix agent failures without reverse-engineering the codebase. ## Core Features & Use Cases - Experiment Architecture Guidance: Explains the experiments/ directory layout (bbbp classification, solubility regression), baseline.json metric configuration, and direction-aware score comparison via is_better() and compute_improvement_pct(). - Git Worktree Workflow: Documents how each run creates an isolated worktree under .worktrees/<experiment>/run_XXX/ with branch naming run/<experiment>/<number>, and how to promote or discard runs. - Debugging Support: Covers common failures such as FileReadTool errors, metric hardcoding mistakes, and per-experiment run numbering issues. - Use Case: You want to run 10 optimization iterations on the BBB penetration experiment and promote the best result as the new baseline; this Skill tells you the exact commands, environment variables (PHARMA_EXPERIMENT, MAX_ITERATIONS), and verification steps. ## Quick Start Ask the assistant to explain how to start an optimization run for the bbbp experiment and how to promote a successful run as the new baseline.
