J.A.W.I.R: Dual-Agent AI Data Governance

J.A.W.I.R (Joint Agent Workflow Integrity & Reliability) tackles a problem that is getting worse as AI coding assistants spread through data teams: routine ETL script changes (a dropped column, a silently retyped field, a renamed feature) quietly break downstream dashboards and machine learning pipelines long before anyone notices. Meanwhile, AI-assisted PR review time has been reported to grow by as much as 91% because reviewers can no longer tell why an AI-generated change was made. J.A.W.I.R solves this with a self-policing Actor-Critic architecture. The solution was built by heavily leveraging IBM Bob 2.0 IDE, utilizing its parallel tasks and subagents to rapidly iterate on complex prompts and debug the UI, while the core runtime engine is powered by IBM watsonx.ai. ASEP (the Actor) drafts pandas ETL scripts and a PR-style justification for any modification request; CECEP (the Critic) independently audits that draft against a schema contract and a data lineage map, rejecting destructive operations (column drops, row filtering, silent type casts, renames) before they ever run, and approving safe changes for one-click execution and export. Built for data engineering teams who manage production ETL pipelines in schema-sensitive sectors like banking, J.A.W.I.R shifts governance left: instead of catching schema drift after a dashboard breaks, it catches it inside the developer's own workflow, in seconds. Internal testing against 5 high-risk modification scenarios showed a 100% detection and block rate. Unlike static downstream tools like dbt, J.A.W.I.R acts as a pre-execution safety gate, giving teams a transparent, auditable reasoning trail for every AI-assisted data transformation.