Agent Bricks: From Experimental Toolkit to Full AI Agent Platform
Databricks Agent Bricks is an AI agent platform for enterprise developers that focuses on model choice flexibility, rich context integration, and strong agent orchestration control so teams can build and deploy high-quality autonomous systems on governed data infrastructure. Launched at a previous Data + AI Summit and expanded at this year’s event, Agent Bricks has already been used to build more than 100,000 agents and now processes over one quadrillion tokens per year in agent workloads. Customers such as AstraZeneca, 7-Eleven, Fox Corporation, Block, Merck, First American and Edmunds have adopted the platform. Databricks is positioning Agent Bricks as the answer to the “missing 99%” of work in enterprise AI development, addressing token capacity, deployment, security, evaluation, monitoring and governance so developers can spend less time on plumbing and more time on application logic.
Model Choice Flexibility Instead of Vendor Lock-In
A central promise of Agent Bricks is model choice flexibility, rather than binding enterprises to a single AI provider. Databricks offers access to major proprietary and open-source models in one security boundary, including OpenAI, Anthropic, Gemini, Qwen and new support for Kimi, plus a partnership to make Grok models from SpaceX available on the platform. Developers can mix models to tune quality, latency and cost for each subagent, instead of overfitting everything to a single large model. Custom models are first-class citizens as well: through Databricks AI Runtime, teams can fine-tune or reinforce-train models on their own data, from small task-specific models to large agent backbones. This approach positions Agent Bricks as an alternative to monolithic AI agent platforms that impose rigid model constraints and make switching providers costly and complex.
Context as a First-Class Primitive in Enterprise AI Development
Databricks frames context as one of the three critical challenges an AI agent platform must solve, alongside choice and control. In production settings, agents need to work across messy data estates that span outdated tables, unstructured file stores, confusing web pages and documents polluted by low-quality AI-generated text. The relevant context may live partly in scattered systems and partly in the heads of a few experts. Agent Bricks extends beyond simple retrieval-augmented generation by supporting tools that search, retrieve and manipulate diverse data during reasoning, all on the Databricks data and governance layer. By tying agent behavior to governed tables, documents, memory and reasoning traces, the platform aims to improve accuracy and traceability. This data-centric design reinforces Databricks’ long-standing claim that connecting data and AI in one environment is the most reliable path to production-grade enterprise AI development.
Control, Governance and Agent Orchestration at Scale
Control is the third pillar of Agent Bricks, reflecting growing concerns around autonomous systems that have powerful tools and access to sensitive data. Databricks highlights examples of agents that have deleted codebases or been prompt-injected into leaking information, and notes that uncontrolled usage can drive token costs sharply upward as developers “tokenmax” benchmarking leaderboards. Agent Bricks responds with secure sandboxes, governed access to data, monitoring, evaluation and quota controls so teams can manage agent orchestration control and cost. It integrates with Databricks Apps to deploy agents with horizontal autoscaling and provides support for multiple agent harnesses, plus the managed Omnigent meta-harness to orchestrate them. This stack is designed to appeal to enterprises that need both flexibility and strong governance, and to differentiate Databricks from agent platforms that prioritize end-to-end convenience over fine-grained operational control.
An Enterprise-Centric Alternative in a Crowded AI Agent Platform Market
The expansion of Agent Bricks comes as the Data + AI Summit draws more than 30,000 data and AI practitioners in person and tens of thousands virtually, with agents and AI as core themes across hundreds of sessions. That audience underlines who Databricks is targeting: enterprise developers, ML engineers and data teams that are already invested in the Databricks Lakehouse and want AI agent capabilities without abandoning existing governance and data pipelines. By emphasizing model choice flexibility, context-aware tooling and rigorous control, Databricks positions Agent Bricks against monolithic AI agent platforms that tie customers to one model stack and proprietary orchestration. According to Databricks, developers have been “stuck building infrastructure, not agents”; Agent Bricks aims to flip that balance, offering a path where infrastructure is handled by the platform so teams can focus on reliable, controllable agent behavior.
