What Agent Bricks Is and Why It Matters Now
Agent Bricks is an AI agent platform from Databricks that lets developers build, deploy, and manage production-grade agents with flexible model choices, rich context access, and strong operational control across the entire data lifecycle. Launched at a previous Data + AI Summit and expanded at the latest event, Agent Bricks now underpins over 100,000 agents and handles more than 1 quadrillion tokens per year, with enterprises like AstraZeneca, 7‑Eleven, Fox Corporation, and Block already in production. Databricks frames this evolution as an answer to a common problem: the core agent loop is only 1% of the work, while the “missing 99%” lies in token capacity, deployment, security, monitoring, and sharing. By turning these hard problems into shared infrastructure, Databricks positions Agent Bricks as a foundation for serious enterprise agent development, rather than yet another experimental framework.
Model Choice and AI Runtime: Flexibility as a First-Class Feature
Databricks centers AI model flexibility as a defining trait of Agent Bricks, arguing that modern agents depend on a mix of subagents and models to balance quality, latency, and cost. On a single secure platform, developers can switch between frontier proprietary and open-source models from providers including OpenAI, Anthropic, Gemini, Qwen, Kimi, and Grok through a partnership with SpaceX. This supports the growing need for model diversity as different releases overtake each other in specific tasks. The platform also builds on Databricks’ AI Runtime for custom models, giving enterprises a path to prompt-optimized, fine-tuned, or reinforcement learning–based models specialized on their own data. According to Databricks, a custom data agent trained with reinforcement learning reached quality competitive with Opus and Sonnet on internal Genie benchmarks while being lower cost per query, giving teams more options than a single default LLM.
From RAG to Rich Context: Reasoning Over the Full Data Estate
In Agent Bricks, context management moves far beyond basic retrieval-augmented generation. Databricks treats context as a core capability, recognizing that enterprise data lives in outdated warehouse tables, unorganized document stores, scattered web pages, and even the memories of a few key people. Agents must search, retrieve, and manipulate data during reasoning, while coping with misleading information and growing volumes of AI-generated “slop” that pollute the estate. Agent Bricks provides the building blocks for this: unified access to data wherever it lives, agent memory, and ample token capacity so agents can work with long or multi-step contexts. Because Databricks already unifies data and AI workloads, agents’ own outputs—reasoning traces, actions, and memories—can be governed, logged, and analyzed alongside traditional datasets. This positions context not as a bolt-on RAG layer, but as an integrated part of enterprise data intelligence.
Control, Safety, and the ‘Missing 99%’ of Agent Operations
Databricks emphasizes that AI agents are among the most privileged actors in an organization, with access to sensitive systems and data. Agent Bricks responds by putting control and safety at the center of the platform. Databricks calls out well-publicized failures—agents deleting codebases, succumbing to prompt injection, and driving costs up through uncontrolled “tokenmaxing”—as symptoms of missing governance. In response, Agent Bricks offers secure sandboxes, governed model access, deployment pipelines, and monitoring to help teams define clear guardrails. Developers can deploy agents to Databricks Apps with horizontal autoscaling, while policy and observability tools keep usage and spending within acceptable limits. Edmunds’ VP of Technology Gregory Rokita summarizes the appeal: “Databricks gives us a secure, governed foundation to run multiple models and switch providers as our needs evolve. All while keeping costs in check.”
Integrating with Agent Harnesses and the Data + AI Summit Ecosystem
Agent Bricks is designed not as a closed stack but as an AI agent platform that works with the tools developers already use. Databricks supports popular open-source agent harnesses such as LangGraph, Agno, and CrewAI, along with SDK-based ecosystems like Claude Code and OpenAI’s Agent SDKs. For teams that want a single orchestration layer, Databricks offers a managed version of its open-source meta-harness Omnigent to coordinate different frameworks. This developer-first stance fits the broader direction of the Data + AI Summit, where more than 800 breakout sessions now span agents, applications, and AI across roles from data engineers to app developers. As Databricks’ founders and industry leaders use the Summit to highlight agents as a central pattern for data and AI, Agent Bricks becomes the practical bridge: a platform that lets teams experiment with any harness, keep their model options open, and still retain enterprise-grade control.
