Databricks open-sources Omnigent, a meta-harness for AI agents, and outlines vision for Agent Clouds

Databricks cofounders Matei Zaharia and Reynold Xin described the company’s strategic pivot toward building infrastructure for enterprise Agent Clouds, arguing that the next durable advantage in AI will come from systems that provide proprietary context, governance, and operational state to agents rather than from frontier model performance alone.

Zaharia and Xin emphasized that coding agents and enterprise agents face shared challenges, including portability, collaboration, session history, security, and spend controls. To address these, Databricks open-sourced Omnigent, a meta-harness designed to sit above existing agent frameworks—such as Claude Code, Codex, Cursor, and Pi—and enable composition, live collaboration, and rich control policies across heterogeneous agent environments.

The executives framed Omnigent as a response to the lack of a common API for agent sessions, tool calls, and cancellation, and highlighted persistent sessions, cloud sandboxes, and sharing as critical features for production-grade agent systems. Zaharia noted that Databricks’ internal usage of agents has scaled to tens of millions of virtual machines daily, underscoring the operational demands of such infrastructure.

Beyond agent orchestration, Reynold Xin discussed Databricks’ work on LTAP and Lakebase, architectures aimed at rethinking the database stack for the agent era. Xin argued that traditional CDC (change data capture) pipelines are brittle and that HTAP (hybrid transactional/analytical processing) has long been an elusive goal. Databricks’ LTAP approach unifies storage layers to provide live operational context to agents, while Lakebase leverages object stores to enable true OLTP and OLAP convergence.

Zaharia and Xin also stressed the importance of governance and security for agent deployments, warning that agents could inadvertently read confidential documents, install compromised packages, or incur significant costs without proper controls. They positioned Databricks’ infrastructure—spanning exabytes of data and tens of millions of daily virtual machines—as uniquely suited to address these challenges at scale.

The executives tied their vision to a broader thesis: as frontier model performance becomes commoditized, the competitive moat shifts to the systems layer that delivers proprietary data, governed access, and operational state to agents. This framing aligns with Databricks’ broader push to evolve from a lakehouse company into what they describe as a data-and-AI operating system for enterprise agents.