Databricks outlines a three-part foundation for scaling enterprise agents

More capable models and coding agents are making agents faster to build and iterate, but operating many agents across an enterprise remains a different problem. [1] As agents shift from answering questions to taking actions, they increasingly depend on models, enterprise data, business semantics, tools, and applications. [1] One workflow may retrieve governed data, choose a model, call several tools, hand work to another agent, update a business system, apply permissions, and leave a trace of what happened. [1]

When teams connect these components independently, Databricks says AI sprawl can produce duplicated integrations, inconsistent policies, higher AI spending, fragmented context, and applications that are harder to change as agent counts rise. [1] Databricks argues enterprise infrastructure should provide choice across models, tools, and frameworks; governed context and business meaning; and consistent control over permissions, policies, evaluation, observability, and costs. [1] Its proposed foundation combines Agent Bricks for building, governing, and optimizing agent fleets, Omnigent as a common layer across agent harnesses, and Unity Gateway for centralized access, cost controls, and observability. [1]

The source separates agent development from fleet operations: easier construction does not remove the need for shared permissions, context, evaluation, observability, and cost management. Databricks is therefore positioning Agent Bricks, Omnigent, and Unity Gateway as a coordinated foundation rather than three isolated tools, with each product assigned a distinct layer of the operating problem. For enterprise teams, the practical test is whether that shared foundation reduces duplicated integrations while preserving choice across changing models, tools, and frameworks. What the sources don't address: how this Databricks stack performs against alternative agent platforms or what measurable cost and reliability gains customers have achieved.