Hervé Bourzeix & Daniel McDonald

At our adtech platform, troubleshooting an ad campaign used to mean opening various tools, joining multiple datasets in your head, and hunting down the one colleague who knew why the campaign isn't performing as expected. We wanted to give that "colleague" to every campaign manager, but we quickly hit a wall: the people who know how to troubleshoot are not software engineers. To solve this, we stopped writing hard-coded agents laden with business logic and let our domain experts become the authors, using a decision-tree-diagram-first approach backed by rules and contextual knowledge. The resulting system uses a prompt management layer that lets experts draft, preview, refine, and publish their specialised agents, which our Campaign Manager can use seamlessly directly from our buying interface. Under the hood, an orchestrator agent walks these Mermaid decision trees and hands off tasks to product-specialized sub-agents, which don't need to know what data exists or in what shape, only what question to ask, through a in-house tool that turns natural language into SQL. This translation engine (NL2SQL) behind the scene is built for production safety: an LLM generates SQL constrained to tenant-scoped Table-Valued Functions (TVFs), having every query checked by an AST parser, validated by a BigQuery dry-run, and finally executed. At each stage, a bounded retry loop feeds errors back into the prompt for self-correction. In this talk, we'll break down the orchestrator/sub-agent topology and explain why it changed our internal contribution model. We'll also dive deep into the security layer, showing how SQL AST validation and TVF-based governance make NL2SQL safe for multi-tenant environments. Finally, you'll leave with a concrete, reusable pattern for building expert-authored systems where engineers are no longer the bottleneck for domain logic.