Data governance typically degrades as data flows through pipelines. Raw tables get joined, filtered, and reshaped into views, and the original context—what a column means, whether it is PII, what quality bar it meets—gets left behind. The result is a few well-governed gold tables surrounded by progressively less documented, less trusted downstream assets. The Governance Agent project, built on Google Cloud Knowledge Catalog, BigQuery, and column-level lineage, takes a different approach: instead of asking humans to re-derive metadata for every downstream view, it traces lineage and propagates governance metadata that already exists upstream.

The agent handles four tasks. Descriptions: if an upstream column has a clear description, the agent traces it through hops and proposes the same description downstream. When a column is transformed (SUM, CASE WHEN, COALESCE), the agent reads the actual SQL and writes a description reflecting the transformation. Business glossary terms: using semantic similarity and unstructured documents (PDFs, product specs, markdown design docs), it maps columns to a controlled glossary rather than guessing from column names alone. Policy tags: the highest-risk task, it traces where PII flows and recommends the same tag downstream, with a summary of who has read access and masking rules. It distinguishes straight pulls from aggregated or anonymized data so it does not blindly stamp PII tags where risk no longer exists. Trust and data quality scores: the agent derives a trust score from upstream Data Quality and Profiling results, giving credit for transformations that improve quality, such as deduplication or null handling. Every propagation runs through a confidence threshold; weak evidence means no automatic action. The design explicitly avoids inferring PII or glossary mappings without solid grounding.

Proactive governance here means the work happens as data moves, not after a scheduled review or compliance incident. New views inherit context automatically, sensitive data is flagged as it flows, and stewards spend time on judgment calls rather than repetitive column-by-column tagging. None of this replaces a data steward; it shifts effort from typing descriptions to deciding what counts as a business term or whether an edge case needs a policy exception.

Lineage is not complete, however. Tables may predate lineage tracking, be one-off imports, or be built by jobs that do not record lineage. For these, the agent accepts your own documents as grounding: short documents can be injected directly into the prompt; long documents are chunked, embedded, and retrieved per column; or an existing Vertex AI Search repository can be queried. The grounding is deliberately conservative. The model is instructed to say when a column is not clearly defined, and to stop—not guess. For policy tags and glossary terms, a column is marked PII only if a document explicitly says so, with no inference from column names.

Lineage is the primary signal, but the agent also uses a second pass. The Data Lineage API only knows what a job recorded; gaps exist. The agent can trigger a Knowledge Catalog Data Documentation scan, an AI-driven analysis by Gemini that infers relationships and column meaning even without a clean SQL trail. Inferred relationships are extracted, cached locally, and loaded into the same traversal engine, so both signals are checked together. Standard lineage runs first because it is grounded in recorded jobs; the Insights pass runs second and fills gaps, with contributions explicitly tagged as inferred. An end-to-end flow handles triggering the scan, waiting for it, extracting results, and applying them, so incomplete pipelines get useful propagation on day one.

Day-to-day operation happens through a Gradio dashboard and a CLI. The dashboard suits stewards previewing tables and approving proposed metadata with a click; the CLI lets platform teams script steward_cli scan, apply, and policy-propagate commands into nightly or CI/CD jobs. Both interfaces matter: UI-only tooling never gets automated, and CLI-only tooling never gets adopted by people closest to the data.

The system is not set-and-forget. Lineage confidence scoring can be wrong across renamed columns or unusual joins; semantic mismatch checks are heuristics, not guarantees. Every propagation is designed to be previewed before applying—that preview is the safety mechanism. The honest pitch is not governance without effort, but governance where effort goes to the right five percent of decisions instead of the repetitive ninety-five percent.

A testimonial from Radina-Paola Ivanova, GenAI Engineer at the VodafoneThree UK Datahub, reports that a significant portion of their data estate was undocumented or inconsistently labelled. By combining cataloguing, lineage, and automated metadata propagation, the Data Steward Agent lets them govern source datasets and scale knowledge across the platform. They estimate cataloguing effort could be reduced by up to 75% while improving discoverability, trust, and AI readiness. The takeaway: governance debt compounds quietly, and lineage-based propagation moves the work upstream so context and controls travel with the data instead of being reconstructed from scratch.