Agents need trusted metadata, so governance is moving into the data layer
Every modern catalog now ships an MCP server, Databricks and Snowflake are sharing governed tables without copying files, and DLP is learning to read the intent behind a prompt.
- DataHub's ranking notes every modern catalog in 2026 ships search, lineage, a business glossary and an MCP server, and argues metadata quality, not feature count, is what separates a catalog that supports AI from one that undermines it. m:…4966
- Databricks says the Iceberg REST Catalog API is recommended for internal sharing of Unity Catalog tables with Snowflake, OpenSharing for cross-domain distribution, and UniForm lets Snowflake read Delta tables directly. m:…2732
- Concentric AI says its Semantic DLP Agent analyses the intent behind AI prompts to catch sensitive data use that conflicts with a user's role, and supports Copilot, Claude, ChatGPT, Gemini and Perplexity plus coding agents including Claude Code, Cursor, Codex and GitHub Copilot. Self-reported m:…3643
- Harvey says it has expanded information governance with self-service retention policies at workspace, vault or matter level, audit logs in the Command Center, and legal holds in early access from early December. Self-reported m:…2223
The catalog is now an agent tool
DataHub's ranking of data catalog tools for 2026 makes a claim that would have sounded odd two years ago: every modern catalog now ships search, lineage, a business glossary and a Model Context Protocol server. The catalog is no longer a website people browse. It is a tool agents call, and the metadata it returns shapes what the agent does next. The ranking argues that metadata quality, not feature count, is the difference between a catalog that supports AI and one that undermines it. m:…4966
DataHub says its own cloud product connects to more than 150 data and AI systems and achieves 97 to 99 percent accuracy on column-level lineage using an open-source, schema-aware SQL parser, and its ranking puts DataHub Cloud first for cross-platform estates that need current, accurate metadata for people and agents. Those are the company's own claims, and the ranking is its own, so they should be read that way. The underlying point stands on its own: an agent that reads a wrong column definition will act on it confidently, and lineage accuracy is now a safety property, not a nice-to-have. m:…4966
Sharing tables without copying them
Databricks describes two ways for Snowflake to read Unity Catalog-governed Iceberg tables without duplicating files. The Iceberg REST Catalog API is recommended for internal, same-domain sharing, while OpenSharing, which Databricks built as an open-source data sharing protocol, is recommended for external, cross-domain distribution. UniForm allows Snowflake to read Delta tables written by Databricks. The governance stays in Unity Catalog while the other platform queries the data where it sits. m:…2732
Gartner estimates annual losses of 12.9 to 15 million dollars due to data inefficiencies, according to the Databricks guide, which is one reason the sharing question matters beyond engineering convenience. DataHub's ranking notes that Collibra and Informatica lead for governance-first, steward-centric programs, and that Alation's AI Governance product registers agents and models from six platforms. The market is splitting between catalogs built for stewards and catalogs built for agents, and the winners will be the ones that serve both. m:…2732 m:…4966
DLP learns to read intent
Concentric AI says it has introduced a Semantic DLP Agent that extends data loss prevention to the endpoint, for desktop and browser-based AI applications. The company says the platform analyses the intent behind AI prompts to detect sensitive data use that is inconsistent with a user's organisational role. That is a change from matching keywords in a document to judging what the user is trying to do with the data, which is the only check that works when the data leaves through a chat box rather than a file download. m:…3643
Concentric AI says the platform supports local AI agents including Copilot, Claude, ChatGPT, Gemini and Perplexity, and coding agents including Claude Code, Cursor, Codex and GitHub Copilot, giving visibility and control over data at rest, data in motion and all AI use in a single platform. The company says Southern Nevada Health District uses the tool to monitor shadow AI and establish guardrails. Shadow AI is the right phrase: the data is leaving through tools the security team never approved, and the DLP layer has to follow it there. m:…3643
Retention and audit catch up with agents
Harvey says it has expanded its information governance capabilities with retention policies, audit logs and legal holds. Workspace administrators can now set retention policies at the workspace, vault or matter level, and retention policy management is now available directly in the Command Center. Retention is the unglamorous half of governance: deciding what an agent's work product is kept for, and what gets deleted on schedule, before a regulator or a court asks. m:…2223
Harvey says audit logs are now viewable, filterable, searchable and exportable in the Command Center, and remain available through the API for integration with security tools like Splunk. Legal holds, which suspend routine deletion of existing data for a custodian, will be available in early access starting in early December, according to the company. Taken together, the four moves in this story point the same way: the controls that used to sit in the application are moving into the data layer, because that is where agents actually touch the data. m:…2223
What is still unresolved
- Whether catalog lineage accuracy claims hold up on real customer estates, not vendor tests
- How OpenSharing adoption compares with the Iceberg REST Catalog API in practice
- Whether intent-based DLP produces false positives that slow legitimate AI use
- How Harvey's legal holds will work once they leave early access
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.