Descriptions:
Latent Space hosts Databricks co-founders Matei Zaharia and Reynold Xin at the Databricks Data AI Summit — an event that has grown from a 50-person Berkeley meetup to 100,000 global attendees, 30,000 in person — for a wide-ranging conversation on the company’s bet that data infrastructure and AI orchestration are converging into what they call the “agent cloud.”
The centerpiece announcement is Omnigents, Databricks’ meta-orchestration layer for AI coding agents. Zaharia explains that internal usage drove the need: advanced engineers were building their own multi-agent workflows and custom UIs on top of even their internal Isaac tool (a Claude Code and Codex wrapper), and customer agents kept hitting the same friction of swapping models and harnesses. Omnigents abstracts across backends so teams can switch underlying models without rebuilding pipelines, while adding a management and analytics plane for tracking agent quality and cost. Dream Engine is covered as Databricks’ initiative to bridge OLTP and OLAP database architectures for AI workloads — explained through a clear primer on why transactional and analytical databases evolved separately and why that split creates problems for AI systems that need both. The episode also touches on Genie (a data science agent), the Panther acquisition, Unity AI Gateway, and Customer League.
Zaharia’s broader thesis: once data is in the right place, generic AGI-level reasoning can be “slapped on top” to produce business value — making data infrastructure quality, not model capability, the primary enterprise bottleneck.
📺 Source: Latent Space · Published June 24, 2026
🏷️ Format: Interview
