Build golden trajectory datasets by curating human-approved agent runs. Use these validated paths to fine-tune agents, create evals, and improve reliability.