AI coding agents have become one of the fastest-growing categories in enterprise software. In the span of just a few years, these development tools have evolved from simple autocomplete assistants into autonomous systems capable of taking over the complete software development cycle, all via natural language prompts.
As vibe-coding takes off, tools from startups like Cursor and Anthropic’s Claude Code have quickly reached multibillion‑dollar revenue run rates. Cursor reportedly crossed $1 billion in annual recurring revenue (ARR) in 2025 and has since approached $2 billion in Q1 of 2026. Anthropic’s Claude Code has scaled even faster, reaching an estimated $2.5 billion annualized run rate within its first year, making it one of the fastest‑growing products in the category that accounts for a large share of Anthropic’s $14 billion ARR.
Yet inside large enterprises, writing code is rarely the hardest part of the job. Data scientists, engineers, and analysts spend much of their time maintaining and augmenting pipelines rather than building new ones. The real bottleneck in enterprise AI, therefore, is not software development itself, but operating complex data systems in production.
Databricks CEO and co-founder Ali Ghodsi believes that the gap represents the next frontier for AI automation. In his view, the next generation of AI agents won’t just write software, but operate the data systems that modern businesses depend on.
That strategic bet is behind Genie Code, a system of autonomous AI agents unveiled today, designed for data engineering, data science, and analytics operations. The system extends the company’s existing Genie platform ecosystem, which allows knowledge workers to ask questions about enterprise data in natural language. (More than 20,000 organizations already used Databricks’s data management and analytics tools; the company’s ARR surpassed $4.8 billion annual revenue in October.)
“Instead of functioning merely as a coding assistant or helping generate code faster, these agents actually understand the structure of the data and existing data problems,” Ali Ghodsi says. “It can automatically set up pipelines, analyze why something is failing, and understand issues like when a dataset schema changes or when permissions are modified.”
For instance, Genie Code can help determine how a dataset should be prepared for modeling—randomizing the data, separating part of it into a test set, or training a model on the remaining portion. After training, the system can aid in evaluating the results using metrics such as F1 scores or the area under the curve, and then analyzing them to determine whether the model is performing well or requires improvement.
“It can suggest trying different approaches—maybe retraining the model or generating plots and graphs to visualize performance, and uncover reasoning about what changes might improve the results,” Ghodsi explains. “It’s not about just generating random code snippets, but understanding the entire structure of the data problem and working through the modeling workflow the same way a data scientist or engineer would.”
Databricks and Enterprise Context
A major reason many AI coding agents struggle in enterprise data environments is context. Most developer tools train primarily on public code repositories and general programming examples. Enterprise data systems, however, add another layer of complexity. Data carries business semantics, governance rules, and access policies that determine how information can be used. Without that context, an AI agent may generate technically correct code that fails once deployed in production.
Genie Code attempts to address that problem by integrating directly with Unity Catalog, Databricks’ governance framework for enterprise data. This integration allows the system to understand data lineage, access permissions, and organizational policies across an enterprise’s entire data estate.
“Maintaining pipelines and making sure they are reliable and always running is a big part of a data engineer’s job, and this is where Genie Code can augment them significantly,” Ghodsi says. “It can monitor systems continuously and respond immediately when something breaks, even in the middle of the night, analyzing complex traces and diagnosing what happened so that the pipeline can be fixed and kept running reliably.”
The architecture relies on a multi-agent architecture powered by multiple AI models. Ghodsi explains that the system combines LLMs from providers including Anthropic, OpenAI, and Google, alongside smaller open-source models optimized for specific tasks. “There are many things inside a workflow where you don’t need a huge model—you just need something fast that can perform a very specific operation reliably.”
The larger models provide the reasoning capabilities necessary for complex problem-solving and planning. Smaller open-source models are trained to handle more routine operations quickly and efficiently. Moreover, the architecture is built around multiple collaborating agents rather than a single monolithic AI system. Each agent specializes in particular functions, such as diagnosing pipeline failures or analyzing data patterns. These agents share context, memory, and skills, allowing them to coordinate their actions and execute complex workflows across the data stack.
Databricks describes this approach as “agentic data work.” Rather than prompting an AI assistant for small pieces of code, users can delegate entire objectives to the system.
Another challenge with autonomous AI systems is maintaining reliable performance in production environments over time, as agents often encounter unfamiliar scenarios that degrade performance. To address that issue, Databricks has acquired Quotient AI, a startup specializing in evaluation and reinforcement learning for AI agents. The company’s technology helps evaluate agent behavior, continuously measuring output quality and detecting regressions before they cause production failures. Quotient AI’s founders previously worked on improving the quality of GitHub Copilot, giving them deep expertise in evaluating AI coding systems.
Vibe-coding for data systems
The rise of vibe-coding has created a new battleground for agentic AI-powered coding tools and reshaped the competitive landscape in software infrastructure. Databricks is approaching the market from a different direction. Ghodsi says the AI coding market and the enterprise data automation market are evolving in parallel but distinct directions.
While tools like Cursor and Anthropic’s coding agents are reshaping how developers write software, Databricks is focused on transforming how companies manage and operate their data systems. “Even though our product name includes ‘code,’ what it really focuses on is data work,” Ghodsi says.
Genie Code targets the workflows that occur after data enters an organization’s platform. By focusing on the data layer, the company aims to address problems that general-purpose coding assistants are not designed to solve. “The other tools in the market help software engineers write application code, which is great,” says Ghodsi, “But for us the end goal is the data: transforming data reliably, and helping organizations work with their data.”
Several organizations, including SiriusXM and Repsol, have already begun experimenting with the technology. SiriusXM uses Genie Code to help build and maintain internal data products, generate SQL queries, and debug pipelines. According to Ghodsi, the company has reported around 20% productivity improvements in data engineering tasks. Genie Code assists engineers in creating data products with defined service-level agreements and reliability guarantees.
Likewise, multinational energy and petrochemical company Repsol is using the technology to accelerate forecasting and production workflows. Instead of manually connecting notebooks, pipelines, and models across different systems, engineers can rely on Genie Code to orchestrate these processes automatically. Ghodsi added that thousands of other customers are already experimenting with the technology, although many deployments are still in early stages.
The Future of Human Engineering
Ghodsi does not expect autonomous agents to replace human engineers. Instead, engineers may spend less time writing code and more time designing architectures, supervising automated systems, and ensuring that AI-driven workflows operate reliably.
“The cost of automation is going down and the tools are becoming easier to use, so naturally the demand for automation increases. If you look at some of the numbers already, a huge percentage of activity on machines is actually agents operating in the background,” he says.
According to the company’s recently released State of AI Agents report, AI agents now create 80% of databases and 97% of test and development environments on the Databricks platform. Just two years ago, agents barely registered in database activity, with human developers handling nearly all of that work.
“I wouldn’t be surprised if that number goes from something like 80% to 99% in a short period of time. But that doesn’t mean humans disappear from the process,” Ghodsi explains. “You also have to think about legal responsibility and quality guarantees. Those are areas where you still need a human in the loop.”