Last year at the Data + AI Summit, we launched Agent Bricks, ushering in a new way to build high quality agents that can reason over your data. Since launch, over 100k+ agents have been built, and we are now processing 1+ quadrillion tokens per year of agents. Customers such as AstraZeneca, 7-Eleven, Fox Corporation, and Block shipped agents built on Agent Bricks. This year at DAIS 2026, we are excited to announce the expansion of Agent Bricks as a comprehensive agent platform for developers.

The missing 99%

The rise of agentic coding, coupled by more powerful frontier models, have unleashed a Cambrian explosion of agents. Building agents with the many agent frameworks or harnesses in the ecosystem has never been easier. However, over the last year, we’ve learned that the core agent loop is just 1% of the work. The other 99% is the hidden technical debt of agentic systems: token capacity, deployment, security, evaluation, monitoring, context, sharing (see below Figure).

Agent platform

Therefore, we observed developers stuck building infrastructure, not agents. This moment calls for an agent platform for developers. 

We believe an agent platform requires solving three critical challenges:

1. Choice. Agents are increasingly composed of many subagents, and need model diversity to strike the right balance of quality and latency. Each model family has unique behaviors and are constantly out-performing each other with every release. Developers need broad model choice, all the way from frontier proprietary and open-source models to cheap but fast smaller models to models customized on their unique enterprise data.
2. Context. LLMs are powerful reasoning machines, but need the ability to retrieve and process the right context in order to make business-correct decisions. This is an extremely difficult problem, as the data estate is littered with missing or misleading information, or the needed context only resides in individuals or needs to be pieced together from multiple sources. 
3. Control. Agents are some of the most privileged actors in an enterprise, with access to sensitive data. The news is replete with agents accidentally deleting codebases, or prompt-injecting to leaking valuable information. And costs are exploding with employees ‘tokenmaxing’ their agentic coding leaderboards. Developers need ways to safely deploy agents, and to control costs so the business can afford to deploy agents at scale.

Building an agent platform that addresses these challenges requires connecting data with AI. After all, agents not only consume data via tools and context, but now also produce lots of data in its output, actions, reasoning traces, and memory – all of which must be governed and analyzed. This unification of data and AI is a feat uniquely positioned for Databricks. 

Agent Bricks

We are beyond excited to announce the next evolution of Agent Bricks as our developer agent platform. What began as an experiment in agent building has expanded into a comprehensive platform for developers to build agents with any model and any harness, access data anywhere, and confidently deploy and control. We have all the building blocks, from secure sandboxes to agent memory to token capacity for developers:Databricks handles the infrastructure while you build impactful agents.

Choice

Models

Agent Bricks offers all the frontier proprietary and open-source models in a single platform, natively integrated into our security boundary. Easily flex and test between different LLMs to balance agent behavior with latency and cost. In addition to OpenAI, Anthropic, Gemini, Qwen, we’ve just added support for Kimi. We’re thrilled to also announce a partnership with SpaceX to make the Grok models natively available on Databricks.

"Databricks gives us a secure, governed foundation to run multiple models and switch providers as our needs evolve. All while keeping costs in check." — Gregory Rokita, VP of Technology, Edmunds

For the last three years, we’ve been pioneering custom models: customers building models specialized on their enterprise data through prompt optimization, fine-tuning, or reinforcement learning. Our research team regularly trains custom models ranging from small models for subagent tasks to applying RL to large models as the core agentic model. Recently, we used reinforcement learning to train a custom data agent that is competitive with frontier models such as Opus and Sonnet in Genie-related tasks, while being significantly lower cost per query (see below Figure). Now, customers such as Merck or First American are using AI Runtime to train LLMs specialized on their unique data.

Agent harnesses

We support any agent harness developers may want to use, from open-source frameworks such as LangGraph, Agno, CrewAI to harnesses such as Claude Code SDK or OpenAI Agent SDKs. Deploy these agents with horizontal autoscaling to Databricks Apps. We also offer a managed version of our open-source meta-harness Omnigent, which we released last weekend, to orchestrate different harnesses.

Context

Retrieving the right data is no longer the RAG applications of yesteryear. Agents now have sophisticated tools to search, retrieve, and manipulate data during reasoning to identify the relevant context. Yet, the demands of today’s agent capabilities require traversing a complex and messy data landscape of outdated tables, unorganized Google Drive folders, confusing web search pages, and misleading documents. Often, the requisite context is simply unrecorded, existing only in the mind of a few key individuals. The rise of AI slop further pollutes the data estate with difficult-to-verify “facts”.

Our research team has been solving critical problems here such as agentic search, memory scaling, programmable scratch pads, evaluation, and grounded reasoning. As part of Agent Bricks, these innovations are delivered in a few key components:

• Connect agents to data everywhere

By adding MCP support to Unity Catalog, agents in Agent Bricks can securely connect to external data sources such as Google Drive, JIRA, Slack, Github, and more. Our specialized search agents are able to leverage both structured metadata and source text to efficiently find the right bits.

• Genie Ontology

By continuously learning an ontology on data, and incorporating human-annotated business semantics, Genie Ontology enables Agent Bricks to access a wealth of information that can guide search and analysis. When does the fiscal year begin? Who is the head of sales? What does a churned customer mean in my business? What is our strategy this year? Which table is most used? What data author has the most authoritative history? Genie Ontology enables agents to instantly understand your business from the start, not have to recreate context with every call.

• Databricks Agent Tools

We’ve shipped a suite of ‘built-in’ tools managed by Databricks that utilize our research innovations to offer best-in-class search of data on the Lakehouse and also external data via MCPs. For example, our agentic search work has yielded a document search subagent that is now 3x faster than before, while improving quality. These tools are centrally accessible and governed in Unity Catalog.

• Agent memory service

Developers building agents can now connect their agents to managed memory on Databricks. Powered by Lakebase under the hood, agents can manage their own context, session history, and persist them across sessions and eventually across agents as well.

• Document intelligence

Ever since our launch last year, a set of functions in SQL we call Document Intelligence (GA) enable state-of-the-art parsing and analysis of PDFs and other documents. With ai_parse_document, ai_extract, and ai_classify, building document processing workflows or subagents is easy. Using our internal benchmark of enterprise document analysis tasks, our system is both highest quality and lowest cost compared to both frontier LLMs and also specialized systems from other providers.

• Databricks Sandbox

Accessing context securely requires careful isolation and access scoping. Databricks Sandbox enables spinning up secure VMs for computing,  downscoped data access to Unity Catalog. These sandboxes can be used to contain code interpreter tools, run subagents and harnesses, or simply as a safe scratchpad for agent experimentation.

Control

The Cambrian explosion of agents, models and tools needs an equally strong counterforce of governance, to safely deploy and manage the cost of these agents. We're thrilled to announce Unity AI Gateway, a unified governance layer across all your AI assets both on Databricks and externally hosted. Every customer should be using Unity AI Gateway to secure, observe, and govern their AI assets, from MCPs to models to external agents.

We have implemented the core capabilities of a governance platform in Unity AI Gateway:

• Discover a catalog of all agents, models, MCPs, Skills, and external agents
• Configure fine-grained access controls for tools and agents
• Monitor cost and enforce per-user and per-group budgets
• Intelligent routing of traffic based on reliability, budget policies, or other controls

But there are a few critical capabilities that only a combined data and AI platform such as Databricks can deliver:

Agent Traces and Monitoring

Agents produce large amounts of data from their reasoning traces, memory writes, and generations. That data should be governed in the Lakehouse alongside the rest of your data, not siloed in a different vendor. The benefits don’t stop there – now that the data is in the lakehouse, apply the full power of Databricks to analyze those traces, to debug agent quality, analyze and optimize AI coding sessions, and monitor behavior in production. Now integrated with LakeWatch, our agentic security platform, configure alerts for PII violations, audit sensitive data access, and respond to security incidents.

Contextual Policies

Agents are stateful, dynamic, and contextual, and so the security policies that govern them should be as well. Build custom security policies for tools, guardrails for agents, directly in SQL (and soon python). Importantly, these policies can hold state and react depending on the data and context. 

For example in the below example, you can write a policy such that, if an agent accesses sensitive customer data with PII, the agent is prevented from publishing that data to a company website, but can email that data to a coworker. Other actions, such as updating Salesforce, would require human approval. 

Unity Catalog Registry for Agents, Tools, and Models

We’ve added agents, tools, and models to Unity Catalog (UC), so you can govern those assets alongside the rest of your data estate. AI governance cannot be separated from data governance. Agents, models, and tools ultimately operate on enterprise data. Governing data and AI together provides consistent policies, end-to-end visibility, and a single control plane for security, compliance, and auditing.

For a comprehensive treatment of AI governance, see the Unity AI Gateway blog.

We are excited to announce Agent Bricks as our fully featured agent platform. We believe that the future of agents requires a combination of data and AI, in a single platform, so that developers can easily build and operate agents in production. By delivering model choice, relevant context, and complete governance, Agent Bricks is ready to build your agentic application. We can’t wait to see what you build.