Preprint reports agent system for document-pipeline tuning

A preprint describes IDP AutoOpt, an autonomous LLM agent for configuring intelligent document-processing pipelines. The authors say the system scores configurations on a small labeled set, diagnoses field-level errors, generates edits and re-evaluates them using human-authored domain skills. In reported extraction, classification and packet-splitting deployments across healthcare, marketing intelligence and financial services, the paper says it cut configuration time from weeks to under two hours. On one extraction benchmark, it reported 90.2% accuracy versus 81.6% for human experts at 4.6 times lower per-page cost.