The catalog is part of the platform now

The catalog’s busiest reader is no longer a person browsing a search page. It is the agent, reading a table’s record before it writes a query. One call returns the whole record: the columns in order with their meaning, the business context, the saved queries others have already proved, the lineage around it, and whether the table can be queried right now.

The context also travels with the results. When the agent runs a query, the answer comes back with the meaning of the columns it used and the tags on them, so a column marked as personal information or deprecated says so at the moment it matters. Text in the catalog reaches the model only after it has been checked, and instructions planted in catalog text are stripped and logged.

Hand-drawn lineage describes how someone believed the data flowed on the day they drew it. Lineage recorded by the platform describes what actually ran. When a query writes a table from other tables, Plexara records that edge and names the audited call that wrote it. When an automation reads tables and exports a file, the edge names the automation. When someone registers a spreadsheet as a table, the edge names the registration.

Because those writes go through the platform, the lineage stays current without anyone maintaining it. The agent uses it too: a column with no description borrows the meaning of the same-named column upstream, up to two hops, and is marked as inherited so nobody mistakes the borrowed meaning for a reviewed one. Searches about a table widen one hop along its lineage, so what the team knows about the source surfaces when someone asks about the result.

The limits are part of the design. Lineage is recorded between datasets, not between individual columns. It covers what runs through the platform. A transformation that runs in a tool outside it is not drawn, and the catalog does not import lineage or metadata from other catalogs.

Four kinds of writer share one catalog. Scans of every connection keep tables and columns current. Tables created on the platform, from uploaded spreadsheets or from webhook sources, are cataloged the moment they exist. People curate in the portal: descriptions, tags, a glossary arranged as a hierarchy, owners, domains, data products, and context documents. Agents capture what they learn while working, and a person reviews it before it is applied.

Every field keeps its own history: who changed it, when, and what it said before. Any change can be rolled back. An edit made from stale information is refused rather than silently overwriting someone else’s work, and a glossary term still in use cannot be deleted out from under the tables that carry it.

What the catalog holds is meaning, not numbers. Agents are told to capture facts about the data, such as what a field means, how a figure is defined, and which table to trust for finance, never the query results themselves. Reviewers strip figures before applying an insight, keeping a number only when it is a stable definition like a divisor or a code. Totals copied into a catalog are out of date by the next quarter, so they are kept out of it.

A catalog bolted onto a warehouse knows about the warehouse. Plexara’s catalog knows about everything the agent can reach through a database, API, or GraphQL connection. Every operation of a connected API is a catalog entry with its schema document, so the agent finds "the endpoint that lists open invoices" the same way it finds the invoices table, and the team can describe and tag it the same way.

In production the catalog holds whole estates. One media organization’s deployment carries thousands of warehouse datasets and more than 200 glossary terms. A retailer’s deployment catalogs the more than 4,000 objects in its ERP beside its point-of-sale data, so a question that crosses the two finds the meaning of both.

The market is moving the same way. Salesforce bought Informatica and closed the deal in November 2025. Fivetran and dbt Labs completed their merger in June 2026 and now describe themselves as data infrastructure for AI agents. Standalone catalogs are being folded into the platforms the work runs on.

Most catalog vendors now offer agent access through the Model Context Protocol, so reaching a catalog from an agent is no longer what sets one apart. What sets one apart is where it lives. A catalog the agent consults on the side is still a second place to keep current, while a catalog inside the platform the agent queries through is updated by the queries themselves.

For a mid-sized company weighing its first catalog purchase, the catalog inside Plexara covers the work that purchase is usually for: knowing what each table and column means, who owns it, where it came from, and which definitions the business agreed on, kept current and readable by every agent and person who needs it.

It is not a full data governance suite. There are no data quality checks or profiling, no data contracts, no column-level lineage, no certification workflow for glossary terms, and no connectors that pull in other catalogs. A company that needs those at enterprise scale will still want a dedicated product. A company that needs its agents and its people to share one accurate picture of its data has it in the platform it already uses.

See Catalog Governance for the product, Plexara vs Data Catalogs for how it compares with a dedicated catalog, and Governing the catalog without leaving the portal for the hands-on lesson. The argument for keeping meaning beside the work is older than this release: Why point-solution catalogs are not enough.