What Is An Agentic Data Catalog?

A traditional data catalog is an inventory of an organisation's tables, dashboards, metrics and owners, written and maintained by people so that other people can look things up. This paper argues that the inventory goes stale for a reason that is arithmetic rather than cultural, and that an agentic data catalog replaces it by changing who does the writing. Agents record the estate as a byproduct of real work, and humans move from authoring entries to approving what the agents propose.

Key takeaways

• •The arithmetic, not discipline: a live estate changes faster than people can write about it, so coverage decays after every documentation push.
• •Agents raise the stakes: a stale description that misleads an analyst costs an afternoon, one that misleads an agent with write access costs an incident.
• •From inventory to control plane: the surface where a person approves, steers or reverses what the agents propose becomes the operational control point.
• •The signature is the product: no agent can make metadata official without approval, and a service account does not count as a name.

Why traditional data catalogs go stale

A platform team stands up a catalog, runs a documentation push, and coverage climbs for a quarter. Then the push ends and the estate keeps moving. Schemas evolve, pipelines are added, metrics are redefined in a transformation layer, tables are deprecated without being dropped. Each of those events invalidates something the catalog asserts. The writing rate is bounded by human attention, and attention goes to whatever is on fire. Documentation is never on fire. So a stale data catalog is not a discipline failure. It is what the curve does when outdated metadata piles up faster than anyone can correct it.

Why an inventory is not enough for AI agents

An inventory answers a person's question, and a person can usually tell when the answer looks wrong. A data catalog for AI agents is read by software that cannot. A fleet of agents also raises the rate of change, because more pipelines get built, more schemas get altered and more access grants get provisioned, so the gap between the estate and the record of it widens at exactly the moment that record starts being read by things with write access. That is why the catalog question is a precondition for agentic data engineering rather than a side quest to pick up afterwards.

From inventory to control plane

The paper sets the two side by side. The catalog as inventory is written by people on a schedule, judged by coverage percentage, and decays to nothing when neglected. The catalog as control plane is written by agents as a byproduct of the work, judged by a short approval queue, and when it is neglected the queue is where work stops. That is the substance of agentic data catalog vs traditional data catalog. An active metadata catalog that projects a governed graph becomes the metadata control plane of an AI-native data platform, the place a person points the agents at a domain and pauses anything mid-flight.

What agents read and write in the catalog

Pages describing tables, metrics, pipelines and owners are generated over the graph the agents write as they work, so nobody documents anything by hand. That is automated metadata discovery and maintenance in its plainest form, because the event that makes a page stale is the same event that regenerates it. Agents also propose the work people used to type, including metadata enrichment and classification of assets. Every claim on a generated page carries a citation that has to resolve to a real governed asset, and a page failing that check is rejected rather than quietly shipped. Natural language data discovery sits on top.

Governance when agents maintain metadata

Human-in-the-loop agent governance here is a property of the path a write travels, not a line in a policy document. An agent-authored fact is stored and readable, plainly marked as the agent's assertion rather than the company's position, until a named person promotes it. A refused promotion is queued for a person rather than returned as an error. A change spanning two clouds arrives as one entry carrying blast radius and evidence, not one ticket per cloud, because the unit of review should be the decision. A long queue is treated as a defect signal, not a throughput metric.

Frequently asked questions

What is an agentic data catalog?

It is a catalog whose entries are written by agents as a byproduct of doing real work, and whose human surface exists for approving, steering and reversing what those agents propose. Pages are projections of a governed graph rather than documents somebody maintains, so there is exactly one place a fact can be true. This is what people are usually reaching for when they ask for an AI data catalog.

How is an agentic data catalog different from a traditional data catalog?

A traditional catalog asks people to author entries on a schedule and measures success by coverage percentage. The agentic version changes the author. Agents write the record as they work, people approve instead of transcribing, and the success signal becomes a short approval queue rather than a coverage number. Its position in the stack changes too, from a place to look things up to the operational control point for the platform.

Why do enterprise data catalogs go stale?

Because the rates do not match. An estate changes continuously as schemas evolve, pipelines are added and metrics are redefined, and each of those events invalidates something the catalog asserts. The writing rate is bounded by human attention, which goes to urgent work, and documentation is never urgent. Coverage therefore peaks during a push and decays afterwards, regardless of how much any particular team cares about it.

Can AI agents maintain metadata without breaking governance?

They can, if the constraint lives in the path a write travels rather than in the agent's instructions, since instructions are exactly what a prompt injection attacks. Agents may assert facts, and those assertions are stored and clearly labelled as theirs. Promotion to an authoritative tier takes a person's signature, and a service account is not a person, so a miscalibrated agent is blocked from poisoning the graph the whole fleet reads.

What is the difference between a data catalog and a context layer?

In the architecture the paper describes, the graph holds the governed facts and the catalog projects them. The context layer is the memory the agents read and write as they work. The catalog is the human surface above it, where the estate is understood and proposals are approved or reversed. The catalog stores no facts of its own, which is what keeps two systems of record from disagreeing with each other.

Do I still need Unity Catalog, Glue or Snowflake Horizon?

The paper does not argue for removing the technical metadata stores each platform ships. Those record what exists inside their own cloud. The argument is about the layer above them: a record of the whole estate written by agents rather than by people, and one surface where a person reviews a cross-cloud change as a single decision instead of one ticket per cloud. That surface is what the paper says is missing.

How does an agentic catalog keep lineage, ownership and quality current?

By computing staleness exactly rather than on a schedule. When something in the graph changes, exactly the pages whose claims depend on it go stale, no more and no less, and the reason is recorded. Generation is also gated on whether the graph supports each stakeholder lens, including owner, governance, reliability, finance, consumer and lineage, because a page that silently leaves a lens uncovered reads as absence rather than ignorance.

How do agents search data by business meaning instead of table names?

Search combines several signals rather than betting on one, fusing semantic search with lexical matching and the shape of the graph, so a weak or slow signal degrades the result instead of hanging it. It classifies what kind of question is being asked before retrieving, and it maps business vocabulary onto technical vocabulary, which is the job a business glossary was meant to do. Churn and cancellation events are the same question.

The whitepaper this is drawn from

This article summarises From the Traditional Data Catalog to the Agentic Data Catalog, a Data Workers Research paper. The full paper carries the figures, the citations and the limitations section, and is free behind a short form.