Agentic Analytics Engineering — The Architecture of Agentic AE

The Architecture of Agentic AE

Stephen Robb (dbt Labs) published the first complete blueprint: dbt + Google AI Agents. Three layers:

1. Local Validation Loop

• dbt compile via Fusion engine validates SQL before warehouse
• Agent generates SQL → validates locally → detects issues → auto-iterates
• Key insight: Fusion's deterministic compiler means AI no longer has to "hope" its SQL is correct

2. Cloud Operations via MCP

• dbt MCP server exposes full project metadata as agent tools
• Agent can inspect models, run tests, understand dependencies — "like an AE would"
• MCP is the critical integration layer: [[AI-Agents-in-Data-Engineering]]

3. Specialised Subagents

• One agent per concern (modeling, documentation, testing, lineage)
• Narrower scope = deeper analysis = more reliable output
• Pattern matches Meta's approach (see [[AI-Agents-in-Data-Engineering]])

Starter repo: github.com/StephenR-DBT/dbt-gemini-agent-starter

Back to [[Agentic-Analytics-Engineering]].