AI Agent

An AI Agent is a first-class catalog entity representing an AI agent — a system built on one or more LLMs that uses tools and skills to accomplish tasks (LangChain agents, Google ADK agents, Snowflake Cortex agents, DataHub's own "Ask DataHub" assistants, etc.). Cataloging agents lets teams discover what agents exist, understand what they can do and what they depend on, govern them, and track their lineage into the data they touch.

Agents come in three flavors, distinguished by the source.type of the aiAgentInfo aspect:

System (SYSTEM): bootstrapped at deployment time (e.g. via YAML) and not editable by end users.
Native (NATIVE): created and managed on DataHub by end users through the UI or API.
External (EXTERNAL): managed outside DataHub but cataloged here for discovery and governance.

This is an origin/governance axis, not a functional one — a single-agent, multi-agent, or copilot distinction (if needed) is layered on with the standard subTypes aspect.

Identity

AI Agents are identified by a single field:

id: A unique identifier for the agent, typically a human-readable slug such as ask-datahub or a generated UUID.

An example URN is urn:li:aiAgent:ask-datahub. The identifier is platform-agnostic; the platform, if any, is carried on the dataPlatformInstance aspect rather than in the key.

Important Capabilities

Agent Info

The core properties live on the aiAgentInfo aspect:

name: Display name, searchable with autocomplete.
tagline: A short (~120 char) one-line summary shown on agent cards, hero subtitles, and picker dropdowns. When absent, surfaces fall back to truncating the description.
description: What the agent does and when to route to it. The recommended style is "X does Y. Use when Z." — for router-style agents (like Ask DataHub) this text is read verbatim when deciding whether to delegate, so the "Use when…" tail is load-bearing rather than decorative.
instructions: Custom base instructions injected into the agent's system prompt, beyond the standard boilerplate.
source: The origin of the agent definition (see below).
created / lastModified: Audit stamps.

Source and Agent Families

The source field (an AIAgentSource) records the agent's origin type (SYSTEM / NATIVE / EXTERNAL) and, optionally, a clonedFrom reference to the agent this one was cloned from (the ClonedFromAgent relationship). Clones inherit their base's configuration verbatim — one frontend module, one tool palette, one set of instructions — and differ only in the artifacts they manage. Grouping a base agent and its clones this way forms an Agent Family.

Dependencies

The aiAgentDependencies aspect captures what the agent relies on, as typed relationships to other first-class entities:

skills → Agent Skill entities the agent adopts (AgentHasSkill).
tools → API entities the agent invokes (AgentUsesTool). This is a lineage edge: the tools are upstream of the agent, contributing to the repo → service → api → agent chain.
models → mlModel entities (the LLMs) the agent runs on (AgentUsesModel).

Versioning, Governance, and Health

AI Agents support the standard governance aspects — ownership, tags, glossary terms, domains, structured properties, and institutional memory — as well as native versioning (versionProperties, so successive revisions of an agent group into a version set) and health via the shared incidents subsystem (incidentsSummary). Agents can also participate in lineage through upstreamLineage.