Wow! The Data and AI Summit 2026 introduced so many monumental announcements for AI in Databricks and none more practical than the recent updates to unity AI Gateway.
Unity Catalog is GA
Let’s start by grounding ourselves on the current (previous now!) state of Unity AI Gateway.
First announced at the previous DAIS summit just two years ago, Mosaic AI Gateway (before the AI Gateway then Unity AI Gateway rebranding) was announced with support for model serving and external model endpoints, along with rate limiting and AI guardrails. Since then the product has matured significantly, and at DAIS 2026 those core capabilities have reached general availability meaning teams can now move beyond experimentation and begin confidently productionising their AI solutions.
With the foundation now GA, the rest of the Summit announcements build on top of it, and the updates are well worth paying attention to.
Introducing Contextual Policies
Traditional governance has always focused on access control ie who can use a model or tool. But as AI agents become more capable and autonomous, that's no longer enough. Organisations now need to control not just who can access a system, but what it can actually do in the moment.
Databricks is addressing this with the launch of Contextual Service Policies in Beta. These policies allow administrators to allow, deny, or require approval for specific actions such as pushing code, modifying files, accessing enterprise systems, or handling sensitive information. Crucially, policies can be applied across multiple dimensions: the user, agent, model, MCP service, tool being invoked, or even the actual contents of a request or response. AI guardrails can also be layered on top to catch risks like PII exposure, prompt injection, jailbreaks, and unsafe content before they become incidents.
In practice, this means an administrator could require human approval before a coding agent pushes anything to GitHub, prevent writes to sensitive Google Drive folders, or automatically block any request or response containing regulated data. For organisations moving toward agentic AI, it means agents can be given meaningful autonomy without sacrificing the oversight and control that compliance and security teams require. This is personally my favourite update from the Summit, as it represents a huge shift in how we think about AI governance, moving from static access rules to dynamic, context-aware controls that keep pace with how AI is actually being used.
For more information on Contextual Policies: Stop rogue AI: How Unity Catalog secures your agent actions | Databricks Blog
Cost Control has Levelled Up
As organisations scale their AI usage, costs have become increasingly difficult to track with token spend scattered across different models, teams, and use cases, making it hard to know where money is going or where AI is actually delivering value.
Databricks is tackling this with four new cost management capabilities in Unity AI Gateway. Unified spend visibility brings together costs from Databricks-hosted models, frontier models, coding agents, and custom applications into a single view, replacing the current reality of chasing spend across multiple systems. Granular cost attribution then goes a level deeper, allowing teams to analyse and budget by user, team, tool, or use case so you can see not just what AI costs, but what it's delivering in return. Hard spend caps provide an automatic safety net, stopping requests when a budget threshold is hit and preventing the runaway costs that can catch teams off guard. Finally, smart routing takes the optimisation further by intelligently directing requests to the most appropriate model based on complexity, quality needs, and cost meaning you're not paying frontier model prices for tasks that don't require them.
Together these features give organisations the visibility and control to treat AI spend like any other managed business cost, rather than an unpredictable variable that grows with adoption.
For more information on Cost Monitoring: AI governance at Data + AI Summit 2026: What’s new with Unity AI Gateway | Databricks Blog
Better Monitoring
As AI systems grow more complex (spanning multiple models, agents, and tools) understanding what actually happened during any given interaction has become a real challenge. Up until now, teams have had to stitch together logs from different systems just to troubleshoot a failure or investigate an incident, which is time-consuming and error-prone.
Databricks is addressing this with three new monitoring capabilities. First, end-to-end agent tracing brings model interactions and MCP tool activity into a single telemetry layer through Unity AI Gateway, giving teams a unified view of how workflows execute across services. Second, Genie can now be used to explore coding agent logs in natural language, making it easier to spot costly or inefficient workflows without having to write complex queries. Third, (and most exciting!), Lakewatch integration allows teams to analyse Unity AI Gateway traces to detect suspicious activity, investigate policy violations, and respond to AI security incidents faster.
Together, these updates mean that instead of piecing together a picture from multiple disconnected systems, teams get a single, governed view of everything their AI workflows are doing, making troubleshooting, auditing, and security investigations significantly more straightforward.
For more information on Lakewatch: The Open Security Lakehouse Built for the AI Era | Databricks
Managed Omnigent
Omnigent was announced as an open source meta-harness for building and running agents across models, frameworks, and coding tools. At DAIS 2026, Databricks introduced the managed version, Omnigent on Databricks, in Beta.
Omnigent on Databricks is the same open source Omnigent, so there is nothing to rebuild. Teams bring their existing setup, harnesses, workflows, and skills, and deploy them to Databricks to run as managed workflows with shared history, remote access, collaboration, and isolated cloud execution on Lakebox.
Unity AI Gateway governs every Omnigent interaction with centrally defined policies, cost controls, smart routing, and unified telemetry. The Summit also brought us teasers for upcoming integrations that will extend support to AI security providers including Alice, CrowdStrike, Cyera, HiddenLayer, Netskope, Noma Security, Obsidian Security, Openlayer, Palo Alto Networks, and Zscaler, as well as identity providers including Okta, Ping Identity, SailPoint, and Saviynt.
This is great news for everyone planning to get started with Omnigent straight away, no rebuilding or infrastructure management and every interaction comes with built-in governance, cost controls, and security, with upcoming integrations extending that coverage to major enterprise security and identity providers.
Learn more about managed Omnigent here: Omnigent on Databricks - Azure Databricks | Microsoft Learn
Final Thoughts
Taken together, these announcements mark a clear turning point for Unity AI Gateway. With the core platform now GA, Databricks has moved the conversation from whether you can govern AI to how precisely you can do it. Contextual Policies bring dynamic, action-level control, the new cost features make AI spend visible and cappable, better monitoring gives teams a single governed view of what their agents are doing, and Managed Omnigent ties it all together. The through-line is clear: as AI grows more autonomous, control has to become more contextual, granular, and unified. For any team looking to move from experimentation to production with confidence, Unity AI Gateway has just become a lot more compelling.
Topics Covered :
Author
Tori Tompkins