Data Agent

Configure Fabric data agents scoped to the open model: generate an agent's configuration from your model, edit its instructions and example queries, preview every change as a dry-run, and publish explicitly. Available in the Studio and over MCP.

#In the Studio

The Data Agent tab lists the data agents in a workspace and lets you open one to see its draft (and published, if any) configuration: the AI instructions, each data source, and the example queries. You can generate a fresh configuration from the open model, edit it, and publish. A persisted auth-mode picker and tenant field sit in the header, so the tab can target the right Entra tenant when your signed-in identity differs from the model's connection.

#Over MCP

Reads are free; every write is dry-run by default.

list_data_agents                       # the agents in a workspace (live read)
get_data_agent <id>                    # one agent's decoded draft + published config
generate_data_agent_config             # build a semantic_model datasource from the open model (Pro)
create_data_agent                      # dry-run by default; commit=true creates
update_data_agent                      # replace only the parts you pass (read-modify-write)
publish_data_agent                     # copy draft to published (dry-run by default)
delete_data_agent                      # dry-run by default; commit=true deletes

#Generate config from the open model

generate_data_agent_config builds a complete semantic-model datasource configuration from the open model in one shot: an element tree of every table, column and measure, carrying the model's descriptions, with each element's selection honoring hidden objects and your Prep-for-AI AI-data-schema exclusions, and the AI instructions seeded from the model's linguistic-schema instructions. It returns the JSON for you to review and writes nothing; feed it to update_data_agent to apply. This is a Pro tool; the free alternative is to assemble the datasource JSON by hand and call update_data_agent. Workspace and artifact ids come back as placeholders, so you resolve the real Fabric ids before applying rather than guessing them.

#Dry-run, then publish

create_data_agent, update_data_agent, publish_data_agent and delete_data_agent are all dry-run by default: with commit=false each returns the exact request it would send and changes nothing, and commit=true executes it. update_data_agent is a read-modify-write that re-emits every existing part plus your changes, so it never drops parts it did not author, and AI instructions are capped at 15,000 characters. publish_data_agent copies the draft parts to published and records a publish description.