Agent Bricks: From Agent Explosion to Enterprise-Ready Platform
Agent Bricks is Databricks’ developer-focused agent platform designed to give enterprises model choice, contextual intelligence, and fine‑grained control so they can move from experimental pilots to production‑scale agentic AI. Announced in expanded form at the Data + AI Summit, it builds on a year of rapid adoption in which customers created more than 100,000 agents and pushed Databricks past 1 quadrillion processed agent tokens per year. That growth exposed a familiar problem: writing an agent loop is the easy part, while token limits, deployment, security, evaluation, monitoring, and sharing form the “missing 99%” of work. Databricks’ answer is to connect agents directly to governed data and unify the infrastructure around them, so development teams can stop building glue code and start treating agents as first‑class applications on a managed enterprise agent platform.
Choice, Context, and Control: The Core of Databricks’ Agent Platform
Databricks frames Agent Bricks around three requirements for serious agentic AI development: choice, context, and control. Choice means broad AI model flexibility. Developers can mix frontier proprietary models with open‑source options, cheaper low‑latency models, and custom models tuned on enterprise data, all within one security boundary. Context focuses on giving large language models precise, business‑correct information across messy data estates, where knowledge is scattered between systems and people. Control addresses the risk side of an agent platform for enterprise use: agents often touch sensitive data and powerful tools, so teams need guardrails against prompt injection, unsafe actions, and runaway token use that can inflate costs. By formalizing these pillars in Agent Bricks, Databricks positions the platform as a way to standardize how organizations build, evaluate, and govern Databricks agents at scale rather than craft one‑off experiments.
Model Flexibility Without Lock-In: Databricks’ Approach to Agentic AI
A central theme of the Agent Bricks launch is freedom from vendor lock‑in for AI agents. The platform brings together leading proprietary models such as OpenAI, Anthropic, Gemini, and Qwen, and adds new options like Kimi and Grok via partnerships, so enterprises can swap or compose models as their needs change. Databricks highlights the growing pattern of agents made up of multiple sub‑agents, each optimized for tasks like planning, retrieval, or code generation. In that world, AI model flexibility is not a nice‑to‑have; it is the core of performance tuning and cost control. As Gregory Rokita, VP of Technology at Edmunds, notes, “Databricks gives us a secure, governed foundation to run multiple models and switch providers as our needs evolve.” That message targets enterprises wary of tying their agent roadmap to a single provider’s pricing, roadmap, or reliability.
Beyond Productivity Pilots: Scaling Agentic AI on the Databricks Stack
Agent Bricks arrives as enterprises try to push AI beyond small productivity pilots into durable, governed applications. Databricks argues that this requires treating agents as part of a full data and AI stack, not as isolated chatbots. The platform plugs into the wider Databricks ecosystem, including Marketplace for sharing data and models, Apps for building interactive experiences, and OpenSharing for controlled distribution across teams and partners. Together, these pieces form an agent platform enterprise developers can standardize on, rather than stitching together separate tools for retrieval, security, and monitoring. Because agents both consume and produce large volumes of data—actions, reasoning traces, and long‑term memory—the integration with Databricks’ governance capabilities is key. It lets organizations apply consistent policies, lineage, and observability to their Databricks agents as they scale them across departments and business processes.
NVIDIA Partnership Powers the Next Wave of Agentic Applications
Underpinning the Agent Bricks vision is an expanded partnership with NVIDIA that targets the computational needs of the “agentic era.” NVIDIA accelerated computing supports training, fine‑tuning, and inference on Databricks AI Runtime, with Hopper GPUs and NVIDIA Quantum InfiniBand networking tuned for large‑scale distributed training and high‑throughput inference. Databricks is also preparing for emerging architectures like NVIDIA Blackwell and highlighting the forthcoming Vera CPU as a foundation for agentic infrastructure. According to NVIDIA’s Pat Lee, the two companies are “supercharging the next wave of enterprise AI by embedding full-stack NVIDIA accelerated computing with Vera CPUs, Rubin GPUs, NVIDIA Quantum InfiniBand networking and NVIDIA Agent Toolkit software into the Databricks platform.” This tight integration means teams building Databricks agents on Agent Bricks can expect consistent performance from experimentation through production, without managing separate hardware silos.
