How are AI agents changing modern data engineering workflows?

One pattern I've found really useful is connecting agents to Fabric through MCP. Instead of the agent guessing, it can look up real run history, table schemas and error messages, then explain what broke and suggest a fix. It feels a lot less like a chatbot and more like a teammate who actually read the logs.

The other big win for us has been documentation. An agent reading notebooks and table schemas from the repo can draft column descriptions and lineage notes in minutes, and someone on the team just reviews them. That's been far more reliable than trying to keep docs updated by hand.

Agree with the points above about keeping it read-only to start. Let it investigate and suggest, and keep the actual changes going through PRs.