When the Codebase Lies: Using AI Agents on a System Nobody Fully Understands

Azure Capacity engineering operates at a scale where no human team can hold the full system in their head. Hundreds of Power BI reports. Dozens of data pipelines. Years of accumulated environment drift between what's live in production and what is in the code repository. Nobody knows which reports are used. Nobody knows which pipelines would break if a schema changed. This talk is about how we used AI agents to make the complexity navigable. We connected MCP servers so agents can see live dashboards eployments and evaluate against our codebase to surface environment drift and usage dead zones, eliminating the reports nobody reads and the pipelines nobody depends on. The talk also focuses on the design decisions: how to structure tool access when the source of truth is split across deployment state and version control, and what incident response looks like when an agent queries the real system rather than documentation which does not reflect the live state. Attendees will leave with concrete patterns for applying agentic workflows to production systems that have accumulated more complexity than any one person can track.
Format: BUSINESS TALK
Location: ROOM A
Speakers: Bogdan Bebic
Level: Beginner
Audience: Registered participants