Descriptions:
Luke Kim, founder and CEO of Spice AI, argues at AI Dev SF 2026 that the modern data stack built for the SaaS era is fundamentally mismatched to the demands of AI agents, and that every agent deserves its own dedicated, isolated data stack. The core problem: agents don’t sleep, operate in continuous loops, and collectively generate orders-of-magnitude more database queries than the human-operated apps they are replacing—something GitHub’s recent outages, driven largely by agentic workloads, make concrete.
Kim’s proposed architecture gives each agent a virtualized data layer—powered by Spice AI’s open-source platform, which can run on a laptop, Raspberry Pi, or as a cloud-hosted service—that sits between the agent and backend databases. This layer handles federated access across OLTP databases, document stores, message buses, and analytical systems, enforcing access policies and firewall rules so agents never touch production systems directly. Local acceleration via embedded columnar databases (including Spice’s own Vortex engine, built on Arrow and DuckDB) keeps the query loop fast without stressing upstream systems.
The session includes a live incident-response demo using an OpenClaw agent connected to Spice’s platform, simulating a production load spike and showing the agent detecting and responding to the degradation autonomously. Kim also references the Lovable security incident as a cautionary example of what happens when agents get unmediated database access, grounding the architectural argument in real-world consequences.
📺 Source: DeepLearningAI · Published May 22, 2026
🏷️ Format: Deep Dive
