Databricks Launches Unity AI Gateway to Govern Every AI Agent You Run

30-Second Summary

Databricks used its annual Data + AI Summit to announce Unity AI Gateway, a platform that puts a single governance, security, and cost-control layer in front of every AI asset an enterprise operates, whether it sits on Databricks or not. Paired with Agent Bricks, the company's new full-stack agent development platform, Databricks is positioning itself as the infrastructure layer for organisations that want to run AI in production without losing control of cost, compliance, or accountability.

At a Glance

• Topic: AI Infrastructure and Governance
• Company: Databricks
• Date: Announced at Data + AI Summit 2026, June 15-18, San Francisco
• Announcement: Unity AI Gateway, a unified governance layer for all enterprise AI assets, and Agent Bricks, a full-featured agent development platform
• What Changed: Enterprises can now govern, monitor, and control every AI model, agent, and MCP server from a single layer inside Databricks, including assets hosted outside the platform
• Why It Matters: The gap between deploying AI and governing AI has been the main brake on enterprise adoption. This removes it for Databricks customers
• Who Should Care: CTOs, IT leaders, and operations teams at companies using Databricks or evaluating enterprise AI infrastructure

Key Facts

• The Data + AI Summit 2026 attracted more than 30,000 data and AI professionals in San Francisco, representing 150 countries
• Unity AI Gateway operates as a centralised runtime registry inside Unity Catalog, applying governance across AI providers, coding agents, agent frameworks, enterprise applications, and custom AI systems
• Four capability pillars were announced: cost controls and smart routing, unified agent tracing, MCP governance, and content filtering guardrails
• Hard spend caps can be set per team, project, or workflow, with smart routing selecting the lowest-cost model that meets the job specification
• Unified agent tracing captures all model and MCP activity in a single governed telemetry layer, with traces available in Lakewatch, Databricks' lakehouse-native security information and event management system
• Real-time rate limits and content filtering are enforced at the gateway level, not inside individual applications
• Agent Bricks combines model choice, relevant data context, and the Unity Catalog governance layer into a single developer platform for building and running agents in production

What Happened

The Databricks Data + AI Summit is the largest data and AI conference in the world, and the 2026 edition ran June 15 to 18 at Moscone Center in San Francisco. On day two, June 16, the company announced two interconnected products that together represent a significant shift in how enterprises approach AI infrastructure.

Unity AI Gateway extends the Unity Catalog governance philosophy to the AI layer. Where Unity Catalog gives organisations a single place to govern data assets, Unity AI Gateway does the same for AI assets. Critically, it is not limited to models and agents running inside Databricks. Any externally hosted model, any third-party coding agent, any MCP server a team has connected, all can be brought under the same governance layer without migrating workloads.

The four announced capabilities address the four failure modes most common in enterprise AI deployments. Smart routing and hard spend caps address runaway cost. Unified agent tracing addresses auditability and debugging. MCP governance addresses the new attack surface created by agents calling external tools. Content filtering addresses compliance and risk management at the generation layer rather than the application layer.

Agent Bricks, announced alongside Unity AI Gateway, provides the development environment where teams build the agents that Unity AI Gateway then governs. The architecture reflects a maturation in how Databricks thinks about the agentic era: build on Agent Bricks, govern with Unity AI Gateway, store and query data in the lakehouse.

Why It Matters

The governance deficit has been the real blocker. Most medium-sized enterprises are not short of tools for building AI agents. They are short of a defensible answer to the question: if an AI agent in your organisation does something wrong, can you explain exactly what it did, why, and how much it cost? Unity AI Gateway is a direct answer to that question.

MCP governance is new and important. The Model Context Protocol has become the standard way agents connect to external tools and data sources. But each MCP connection is also a new data flow, a new cost centre, and a new security surface. Logging and governing MCP traffic at the platform level, rather than trusting each application to do it, is a meaningful upgrade.

Cost predictability unlocks budget approval. One of the most common reasons enterprise AI projects stall is that finance teams cannot approve an open-ended AI budget. Hard spend caps that enforce predictable tokenomics across automated workflows convert AI spending from a variable operational risk into a manageable line item.

Databricks has distribution. Other companies have built governance layers for AI. The difference here is that Databricks sits inside the existing data stack of thousands of large enterprises. Unity AI Gateway does not require a new vendor relationship, a new security review, or a new procurement cycle for organisations already on the platform.

The standard is now set. Enterprises evaluating AI infrastructure vendors now have a clear benchmark: a single governance layer for all AI assets, regardless of where they are hosted. Any vendor that cannot match this is now behind.

The David and Goliath View

The 2026 enterprise AI story is not about which model scores highest on a benchmark. It is about which infrastructure lets a real organisation with real compliance requirements, real budget constraints, and real security obligations run AI in production without gambling on the outcome. The Databricks announcements at DAIS 2026 are the clearest articulation yet of what that infrastructure looks like.

For organisations that are already running AI agents, Unity AI Gateway closes a gap most of them know they have but have not yet fixed. Ungoverned agents running on multiple models with no unified logging, no spend caps, and no MCP visibility are not a future risk. They are a current one. The platform makes it possible to address that without rebuilding the stack.

The pairing of Agent Bricks and Unity AI Gateway is also worth noting as a product strategy. Databricks is not simply offering governance as an add-on. It is offering governance as the foundation, and building the development environment on top of it. That ordering matters. It means governance is not something you retrofit. It is something you build into from day one.

Where This Fits in the AI Stack

Unity AI Gateway sits between the agents and the outside world. Every request an agent makes to a model, every tool call to an MCP server, every external API connection, passes through the gateway before it executes. This positioning makes it a policy enforcement point, a cost management layer, and an audit log simultaneously.

Agent Bricks sits one layer up, giving developers the environment in which to build agents that are, by design, compatible with the governance layer beneath them. Together, they represent the data plane and the control plane of an enterprise AI system built on Databricks.

Questions Operators Are Asking

We already have agents running. Can we add Unity AI Gateway without rebuilding them? Yes. The gateway is positioned as an overlay, not a rebuild. Existing agents, models, and MCP connections can be routed through the gateway without changing the applications themselves. Databricks has designed it to wrap existing infrastructure rather than replace it.

What does MCP governance actually mean in practice? Every time an AI agent calls an external tool via MCP, that call is logged, timed, and attributed to a cost centre. The gateway can enforce which MCP servers are permitted, at what rate, and under what conditions. This gives security teams visibility into a layer of AI activity that has, until now, been essentially invisible.

Does this only work if we are already on Databricks? Unity AI Gateway is designed to cover externally hosted assets, meaning models from OpenAI, Anthropic, Google, and others can be brought under the governance layer even if those workloads are not running on Databricks compute. However, the deeper integration benefits, tracing in Lakewatch, Unity Catalog asset governance, and Agent Bricks integration, require the broader Databricks platform.

How does smart routing decide which model to use? Smart routing applies a set of policy rules defined by the operator: cost thresholds, latency requirements, capability requirements, and any model restrictions the organisation has set. Within those constraints, it selects the lowest-cost model that meets the job specification. Operators can also set hard overrides for specific workflow types.

What is the compliance case for regulated industries? The combination of unified agent tracing, content filtering at the gateway level, and hard spend caps with auditability creates a defensible governance record for regulatory purposes. Financial services, healthcare, and legal teams can point to a centralised log of what every AI agent did, what it cost, and what content filters were applied.

Citable Summary

Databricks announced Unity AI Gateway at its Data + AI Summit 2026, a unified governance layer that covers every AI asset an enterprise operates whether hosted on Databricks or externally, including models from competing providers and external MCP servers. The platform introduces hard spend caps, smart model routing, unified agent tracing, and real-time content filtering from a single control point inside Unity Catalog. Paired with Agent Bricks, the company's new developer platform for building production agents, the announcement sets a new infrastructure standard for governed enterprise AI. The Data + AI Summit 2026 ran June 15 to 18 at Moscone Center in San Francisco, with more than 30,000 attendees.