|
Every team running AI agents has the same quiet problem: the instructions that make an agent useful live everywhere. A prompt in someone’s notebook, a runbook in a repo, a “here’s how we join these tables” message in Slack. Unity Catalog Skills is Databricks’ answer, and the tutorials call it Unity Gateway Skills, because Databricks hasn’t settled on one name. The August 28 release note puts the idea in one sentence: “A skill is a first-class Unity Catalog securable (catalog.schema.skill) that follows the SKILL.md specification.” Publish it once, and you control who can discover, use, and update it with “the same permissions, tags, and audit that protect your tables.”
The format itself isn’t Databricks’. Agent Skills is an open format that Anthropic originally developed: a folder holding a SKILL.md file with a name, a description, and instructions. The spec’s client showcase lists more than 40 agent products, from Claude Code, Cursor, and GitHub Copilot to OpenAI’s Codex, Gemini CLI, Databricks’ own Genie Code, and Snowflake’s Cortex Code. The file format is a commodity. What Databricks is selling is the governance around it.
That governance is the part an architect should care about. You don’t copy a skill to a teammate, you grant it. A reader needs the same three privileges they’d need for any volume: USE CATALOG, USE SCHEMA, and READ VOLUME. The governance page says Unity Catalog “records control-plane operations on skills, including create, update, delete, read, and permission changes,” so the record of who read, changed, or was granted which instructions sits in the same audit as your data access.
|
“Sharing a skill is a grant, not a copy: the recipient’s agent reads the live skill under your grants and audit, so there is nothing to keep in sync.” Databricks docs, Create and share Unity Gateway Skills.
|
Getting an agent to use them takes a small command-line tool. You install Databricks’ Unity Gateway CLI (Python 3.12 or later), point it at your workspace, and your coding agent can then list, load, or download skills over MCP. The tutorials name five agents: Claude, Codex, Gemini, opencode, and Copilot. An agent can load a skill for one session, download a copy to disk, or load a whole schema of skills live.
Who can do what
| Use a skill |
USE CATALOG, USE SCHEMA, READ VOLUME |
| Create one |
USE CATALOG, USE SCHEMA, CREATE VOLUME |
| Update one |
USE CATALOG, USE SCHEMA, WRITE VOLUME + READ VOLUME |
| Share one |
Owner or MANAGE, granting READ VOLUME |
Skills use volume privileges rather than having their own. Source: Databricks’ Govern skills page. The page doesn’t say which privileges delete a skill.
Skills are new enough that the useful questions are about the edges: versioning, Git, who can author, and who is blocked. Here is what the docs say, including the parts the release note leaves out.
Downloaded copies go stale. The docs call a downloaded skill “a point-in-time copy”: re-run the download for a newer one, or load the schema live to always get the latest. Live loading works on a whole schema, so the schema is your unit of freshness. The pages we read describe no version history and no way to pin a skill to a version.
Git sync exists, but it never deletes. Databricks provides a notebook that syncs a Git repository into a schema, and it can run on a schedule. In the docs’ words, it “creates new skills, and updates changed ones; it never deletes, so a skill removed from the repo stays in the schema until you drop it.”
Updating takes two privileges, not one. Authors need WRITE VOLUME and READ VOLUME together, and the governance page says WRITE VOLUME alone isn’t sufficient. The create-and-share tutorial’s prerequisites list only USE SCHEMA and CREATE VOLUME.
Beta, behind an account toggle, on standard catalogs only. An account admin has to enable the Unity AI preview from the account console Previews page first. Skills need a standard Unity Catalog catalog on managed storage, not a foreign or federated one, and Databricks suggests a dedicated skills catalog with one schema per domain or team.
Compliance workspaces can read skills, not publish them. The governance page says that in workspaces enforcing the compliance security profile, “creating, updating, and deleting skills is blocked,” while reading existing skills is still allowed. On September 17, Unity Gateway itself became available by default in workspaces with the profile enabled, but that release note says nothing about skills.
Databricks’ own agent isn’t on the list yet. The tutorials name five third-party coding agents. Genie Code’s skills page, updated September 11, keeps skills in workspace and per-user folders, and its one mention of Unity Catalog is a built-in skill for exploring data, not a place to publish skills.
So: should your team’s agent instructions move into Unity Catalog? If several agents already work against the same data and your instructions are drifting between copies, pilot it now. The permission model is one your platform team already runs, and putting the instructions in the same audit as the data is the strongest reason to do it. It’s Beta, so treat it as a pilot, and keep Git as the source of truth with Unity Catalog as the published registry, because the sync never cleans up and the docs describe no version history. If your workspaces run the compliance security profile, or you were counting on Genie Code, the current docs don’t cover you yet.
|
