Databricks Launches Agent Bricks, Its New No-Code AI Agent Builder

If you’re a tech company of any stature and you haven’t yet released a way to build agents on your platform, what are you even doing with your investors’ money?

Last July, Databricks launched its Mosaic AI Agent Framework, which was one of the  earlier entrants into the AI agent space. The framework’s focus was on using retrieval augmented generation (RAG) to pull in data and documents, but for the most part, it wasn’t all that different from any other agent framework like LangGraph, CrewAI or Autogen in the market today.

Today, Databricks launched Agent Bricks, a very different take on building agents and multi-agent systems on its data platform, at its Data+AI Summit in San Francisco this week. The twist here is that you don’t actually have to build the agents yourself but simply provide the service with a high-level description of what you want them to do, and the system will do the rest (though you can always delve into the code).

Since this is Databricks, the company founded by the team behind the Apache Spark big data analytics engine, the focus for the Agent Bricks service is on agents that can help its users extract data from unstructured information, handle text transformation and create chatbots based on all of this data.

“The hype about agents are everywhere, but agents themselves aren’t everywhere,” Databricks’ VP of Marketing Joel Minnick told me. He noted that many Databricks customers are experimenting with agents and building proofs of concept, but very few are making it into production. “It’s coming down really to two things: cost and quality. If I’m going to put this thing out there in production, then I have to be able to operate at a cost I can sustain, and I have to have a lot of trust that it does what I think it’s supposed to do,” he said.

The Missing Pieces: Evaluation, Data and Optimization

A few pieces are missing here, Minnick argued. Among them is a better methodology for testing these agents. The typical benchmarks being run by the model providers don’t, after all, look at how these models and agents will work in an enterprise environment with very specific guardrails.

Many enterprises also may hoard a lot of data, Minnick noted, but it may not be the right data to train a model that can help with customer service, for example. For them, Agent Bricks will create synthetic data to train their agents.

Then, these companies also have to figure out how to best optimize these models for their use cases — and which models to use. “What we found is that this was just a lot of trial and error for a lot of companies. Like, well, let’s try this and let’s see if that works, and it’s stretching out the amount of time they’re spending trying to build these things into a place that’s very expensive,” Minnick said.

Ideally, this means that, going forward, a Databricks customer could simply tell Agent Bricks to build an agentic system that can answer questions about the products their company offers and the system will go ahead and set that up. The users can set up LLM judges with a set of evaluation criteria and since Databricks has a deep understanding of its customers’ data, it can then create the right synthetic data for this use case to help the agents learn.

From there, Databricks will automatically choose the right model and optimize it based on the user’s constraints. Maybe, Minnick said, the Llama 7B model may get a user to 95% of the quality they want at a given cost, but using a cheaper, optimized model may get them 87% of the quality at a third of that cost.

As those agents get used in production, the system will regularly run through these optimization loops again.

In the background, Agent Bricks builds these agents with the Agent Framework at their core.

One of the companies that Databricks has been piloting this service with is the pharmaceutical giant AstraZeneca. In an hour, Minnick said, their developers built an information extraction agent using Agent Bricks and at this point, that agent has processed over 400,000 documents and is helping researchers and clinicians understand data from clinical trials.

“Agent Bricks is a whole new way of building and deploying AI agents that can reason on your data,” said Databricks CEO and co-founder Ali Ghodsi. “For the first time, businesses can go from idea to production-grade AI on their own data with speed and confidence, with control over quality and cost tradeoffs. No manual tuning, no guesswork and all the security and governance Databricks has to offer. It’s the breakthrough that finally makes enterprise AI agents both practical and powerful.”

In addition to Agent Bricks, Databricks is also launching a number of additional features for its Mosaic AI platform. These include support for serverless GPUs for fine-tuning models, running machine learning workloads and experimenting with large language models, all without having to manage any GPU infrastructure.

The company is also launching a new version of its MLflow platform for managing the AI lifecycle, with new tools for monitoring and tracing agents, no matter where they are hosted.