What Agent Bricks Is and Why Enterprises Care
Databricks Agent Bricks is an AI agent platform that lets enterprises build, deploy, and govern AI agents that reason over business data while giving teams model flexibility, deep context, and operational control in a single environment. Since its launch, Agent Bricks has moved from an experiment into a core enterprise AI agent platform, with over 100,000 agents created and more than one quadrillion tokens processed each year. Early adopters such as AstraZeneca, 7‑Eleven, Fox Corporation, and Block have shipped production agents that draw on governed data while staying inside Databricks’ security boundary. This shift reflects a key lesson from the past year of agentic applications: writing the core agent loop is only a small fraction of the work, while deployment, token capacity, evaluation, monitoring, and security represent the hidden technical debt that blocks enterprise scale.
Model Choice, Context, and Control: The Three Pillars of Databricks Agent Bricks
Agent Bricks is designed around three linked needs for enterprise AI agents: choice, context, and control. On choice, the platform brings frontier proprietary and open‑source models into one place, including OpenAI, Anthropic, Gemini, Qwen, Kimi, and Grok models via a SpaceX partnership, plus support for custom models trained or fine‑tuned on enterprise data. This lets teams mix high‑quality, slower models with cheaper, faster ones across sub‑agents to balance latency and cost. Context is handled by tying agents directly to the enterprise data estate so they can retrieve and assemble the right information for decisions, instead of relying on incomplete prompts. Control comes from treating agents as privileged actors: their access, actions, token use, memory, and traces are governed on the same platform where the data and models live.
NVIDIA AI Infrastructure: The Engine Behind Agentic Applications at Scale
NVIDIA AI infrastructure underpins Databricks’ push toward production‑grade agentic applications. Inside Databricks AI Runtime, NVIDIA Hopper GPUs and NVIDIA Quantum InfiniBand networking support large‑scale training and fine‑tuning directly alongside governed data, removing the need for separate GPU clusters. Databricks is also preparing for NVIDIA Blackwell GPUs, and plans support for NGC containers and custom CUDA environments, so enterprises can standardize on NVIDIA’s stack across the AI lifecycle. On the inference side, Databricks Model Serving uses NVIDIA hardware and Triton Inference Server to provide low‑latency, high‑throughput serving for both frontier and custom models, including Qwen and GPT‑OSS. As Pat Lee of NVIDIA notes, the expanded partnership embeds “full‑stack NVIDIA accelerated computing with Vera CPUs, Rubin GPUs, NVIDIA Quantum InfiniBand networking and NVIDIA Agent Toolkit software into the Databricks platform,” aligning infrastructure with the needs of AI agent platforms.
From Multi-Tool Chaos to a Unified AI Agent Platform
Enterprise teams building agentic applications often juggle separate systems for data access, model serving, orchestration, observability, and governance. Agent Bricks aims to replace this multi‑tool sprawl with a single AI agent platform that combines data, models, and runtime. Because agents both consume and generate large volumes of data—context documents, actions, reasoning traces, and memory—their full lifecycle is managed inside Databricks’ governed environment instead of scattered across services. NVIDIA accelerated computing makes this consolidation practical at scale, powering training, fine‑tuning, and inference for thousands of customers. According to Adam Conway, Databricks’ SVP of Product, the partnership with NVIDIA enables “enterprise AI that’s fast, scalable, and built on a foundation they can trust.” The result is less infrastructure work for developers and a clearer operational model for running critical enterprise AI agents.
Flexibility Without Losing Governance and Security
A central promise of the Databricks and NVIDIA collaboration is that enterprises can choose the right model for each AI agent task without giving up governance or security. Because Agent Bricks integrates many proprietary, open‑source, and custom models inside one security boundary, teams can switch providers or mix models without exposing sensitive data to uncontrolled endpoints. NVIDIA Vera CPUs, Rubin GPUs, and Quantum InfiniBand networking are embedded in the platform rather than managed as separate infrastructure, which keeps compute close to governed data and simplifies compliance. At the same time, controls for deployment, token capacity, and cost management help firms prevent misuse and keep agentic workloads sustainable. In practice, this means enterprises can scale from a handful of experimental agents to thousands of production‑grade enterprise AI agents while keeping ownership of both their data and AI stack.
