Data Systems in the Age of AI Agents: Requirements and Architecture

AI agents don’t think sequentially. They speculate, launching thousands of queries per second, 80% of them redundant. Legacy warehouses weren’t designed for that pattern. The result: latency spikes, runaway costs, and broken context.

Inside you’ll find:

• The 4 defining characteristics of agentic workloads (from Berkeley’s 2024 research)
• Why sub-second performance is non-negotiable for agent iteration loops
• How redundancy creates a 10× compute tax — and how to reclaim it
• Design principles for agent-ready architectures (sub-query reuse, dynamic scaling, standards-based integration)
• Why Firebolt’s architecture aligns naturally with these requirements today

This is the first resource that translates Berkeley’s agentic speculation research into real architectural guidance.

Why now?

Agentic workloads are coming whether you plan for them or not. You can either read this paper, or rediscover its lessons the hard way in production :).