Part VI
D3M with AI Agents
When the analyst is an agent
This closing part surveys the frontier the rest of the book leads to: a data-to-decision loop operated by AI agents rather than by hand. It is a grounded, citation-backed survey of the state of the art in mid-2026 — what is real, what is hype, and where it is headed. It defines what a data agent actually is, then takes the four capabilities that make agentic analytics work in turn: querying production databases in natural language through a governed semantic layer, running automated predictive workflows, connecting agents to the data stack through MCP and orchestration, and the evaluation, security, and governance that keep all of it safe. It closes not on a tidy capstone but on a contradiction — bold forecasts beside sober failure rates — and the argument that the discipline this book teaches matters more in the agentic era, not less.
1 chapter · 6 articles
What you’ll learn
- Tell a genuine agent from a workflow, and read the anatomy and autonomy dial of a data agent
- Judge text-to-SQL honestly — why it is near-solved on clean schemas yet fragile on real ones, and why the semantic layer is the fix
- See how far data-science agents have come on real benchmarks, and where the monitor-and-retrain loop pays off
- Map the agentic stack — tool use, MCP, A2A, orchestration, durable execution — and connect agents to data safely
- Evaluate, observe, secure, and govern a production data agent against the lethal trifecta, NIST, the EU AI Act, and ISO 42001
In this part
What changes when AI agents — not analysts — operate the data-to-decision loop.