2026-09-11 –, Room 1 (170)
As Large Language Models evolve from passive chatbots to autonomous agents, the way they consume data is fundamentally changing. Traditional data platforms are built for human analysts—optimized for static dashboards, batch processing, and ad‑hoc read-only queries. But what happens when your primary data consumer is an autonomous agent that needs real-time context, semantic understanding, and the ability to take action?
This talk bridges traditional data engineering and the emerging needs of agentic AI. We’ll explore the architectural shifts required to build agent-friendly data platforms: robust semantic layers, data access via deterministic tool/function-calling interfaces, and strict guardrails so agents interact with data safely and predictably.
Key takeaways:
The fundamental differences between analytical and agentic data consumption
How to build semantic layers and tool-calling interfaces for autonomous agents
Best practices for auditing, rate-limiting, and securing agentic database interactions
While the AI ecosystem has hyper-focused on model capabilities, the data infrastructure required to support autonomous agents in production is often an afterthought. This talk provides a practical, vendor-neutral blueprint for data engineers supporting internal AI agents.
The technical depth focuses on the shift from SQL-centric, batch-oriented warehouses to API-driven, semantically rich data layers designed for tool-using agents. We’ll address core engineering challenges in agentic use cases:
Semantic translation: agents struggle with raw, normalized schemas. We’ll cover how semantic layers (metrics/dimensions/contracts) give agents deterministic business definitions instead of “best guess” SQL.
Tooling & APIs: moving from direct SQL access to scoped tools (REST/GraphQL, stored procedures, or function-calling endpoints) that agents can use reliably without hallucinating table names or join paths.
State & memory: combining OLAP (analytics) with vector search (context retrieval) and lightweight OLTP/state stores (work tracking, approvals) so agents have both short-term context and durable memory.
Governance (the blast radius): enforcing RBAC for agent service accounts, query/endpoint allowlists, rate limits and budgets, and audit logs that let you answer “what happened, who/what asked, and why?”
This talk avoids AI hype by focusing strictly on the backend engineering and infrastructure required to make agents reliable, safe, and performant.
