Starburst Brings AI Agents to SQL Developers—No Machine Learning PhD Required
The data platform company Starburst is rolling out features that put AI capabilities directly into the hands of SQL developers, eliminating barriers that have forced engineers to rely on specialized machine learning teams or spend months building complex pipelines.
At its AI & Data Nova conference in New York, Starburst announces updates to its data platform that address a fundamental problem: business leaders want automated decision-making and AI-powered insights, but data infrastructure complexity keeps getting in the way.
SQL Developers Get Direct Access to LLMs
The centerpiece of the announcement is a set of AI SQL functions that let developers call large language models directly from SQL queries. Tasks that previously required spinning up Spark clusters and building machine learning pipelines can now be handled with familiar SQL syntax.
"You're a machine learning engineer, and you have received some review data, and your job is to classify data," explained Jitender Aswani, Starburst's SVP of Engineering, during a briefing. "You're doing feature extraction, analyzing sentiment—is this review positive, negative, neutral? Now you can do that in SQL."
The functions include prompt(), classify(), analyze_sentiment(), translate(), fix_grammar(), and mask(). Developers can customize prompts for each function, maintaining control over how the LLM behaves—a critical governance requirement for enterprises.
This matters because SQL is the world's most widely used data language. Starburst estimates there are one billion SQL users. By bringing AI capabilities to SQL rather than forcing developers to learn new tools, the company is "dramatically bringing down all the barriers," Aswani said.
Data Products Replace Pipeline Hell
Starburst is doubling down on what it calls "data products"—packaged datasets that include not just the data itself, but also metadata, business rules, access patterns, and governance controls.
"I've been on both sides of both data producers and data consumers, and they have this classic tug of war," Aswani noted. "They can't talk to each other unless there is a common middle ground, and data product is that common middle ground."
Without data products, AI agents either can't access the data they need or require multiple agents performing precise actions and handing off to other agents—an approach that's expensive, brittle, and low quality.
Aswani compared data products to the connectivity layer that made the internet explode. "You can have all the tools," he said, referring to competitors like Databricks and Snowflake. "But these tools are not connected. They're not using data products. So now it's fragmented."
The company has added agent-assisted data product creation, automating documentation and making it faster to set up governed, reusable data assets.
Conversational Analytics That Actually Work
Starburst's new AI agent lets users ask questions in natural language and get back text insights, visualizations, or SQL code. Users can choose from three personas—executive, analyst, or data engineer—each tailored to different needs.
But executives are "a little surprised they got so excited," according to Nathan Vega, Senior Director of Product Marketing. "They don't need to ask a team to go make a dashboard for me."
Engineers like this too. It keeps them focused on substantial work instead of building one-off dashboards for executives who'll use them for a week and move on.
The agent can be embedded headlessly into any application via a REST API, giving developers a white-label AI capability they can integrate into their own workflows without building from scratch.
Multi-Agent Management Built In
Starburst is implementing the Model Context Protocol (MCP) to enable agent-to-agent communication. The company already has customers running thousands of agents in production.
"A chief data officer looking for a new data platform will have a personal agent that runs, send it off to go get five RFPs from companies," Vega predicted. "That agent will interact with our agent, have it fill out the information and come back to help make a decision."
The platform includes model access management that lets enterprises control which models different users and agents can access, enforce rate limits, and monitor token usage—critical for cost control and compliance in regulated industries like banking and healthcare.
"Governance needs to be very unsexy," Aswani said. "This is the unsexiest part of managing a data platform, but critical."
No More Vendor Lock-In
Unlike competitors that require centralizing all data into their proprietary platforms, Starburst works across hybrid and multi-cloud environments, including air-gapped on-premises deployments. This matters for regulated industries that can't move certain data to the cloud.
The platform runs structured and unstructured data in one place, eliminating the need to wait on pipeline creation. For developers tired of ETL brittleness, this is exactly what "localized innovation" looks like—enabling teams closest to customers to build solutions quickly without depending on centralized data science teams.
Most features will be available in Q4 2024, with some capabilities like the MCP server arriving in early 2025.
