New study suggests major productivity boost when using Cursor's coding agent

Estimated reading time: 3 minutes

The University of Chicago study is the latest indicating AI coding agents are starting to make a real difference.

The potential of AI coding tools to supercharge developer productivity has been much discussed – but not always evidenced. Trust in AI coding tools is plummeting, according to recent surveys, as devs get their hands on them and weigh their benefits and drawbacks. In the industry, there’s an AI competence penalty against those who utilize the tech.

But a new study from a researcher at the University of Chicago suggests that the benefits are real – and potentially enormous. The analysis of tens of thousands of developers across 1,000 organizations suggests that Cursor’s AI coding agent specifically can massively increase software output, all without negatively affecting the fix rates or revert rates.

The study found that companies using Cursor’s agent as their default merge 39% more pull requests than those that don’t. That’s based on analysis of 24 organizations that were using Cursor, compared to eight similar companies that weren’t.

Those who adopted the agent saw weekly code merges rise by more than a quarter. At the same time, there was no significant rise in short-run revert rates, while the sharing of bug fixes dropped, suggesting that the code produced was strong enough to be deployed without issue.

Increasing evidence in favour of AI tools

Alex Fazio, a Python programmer and co-founder of AI Garden, an AI consultancy, came across the study while trawling the web. “I normally read papers mostly because I’m looking for confirmation of what I’m seeing,” he says.

The paper was “one of many that came out recently that’s confirming what I’m seeing personally in my development work: that AI is getting increasingly useful to code, to the point that I think you’re being left behind if you’re not using it.”

Fazio believes that LLMs can, with the right method, be applied across most coding tasks, but they are especially powerful in areas that teams tend to neglect – such as testing, documentation, and clearing long-ignored backlog items.

Although the study doesn’t outline specific tests or documentation in the headline findings, the sample prompts that the study suggests are used by organizations include refactoring large files, adding docs, drafting test plans, and running test suites or build pipelines – the kind of maintenance and clean-up work that often lingers in teams’ backlogs.

The study also found other areas where AI assistance can be applied. An analysis of the first messages sent to the Cursor agent suggested that nearly two in every three of them were asking to implement code, with others asking the agent to explain errors, interpret existing code, or draw up a plan.

More experienced coders were more likely to ask the agent to do these more interpretative or planning processes than actually writing code – something that the author, Suproteem K. Sarkar at the University of Chicago, suggests is an indication that “agents may shift the production process from the syntactic activity of typing code to the semantic activity of instructing and evaluating agents.”

The different use of agents depending on experience highlights how using AI for help is becoming less taboo, says Fazio. “So far, the conversation has been about how it can help junior developers to be effective, which makes it look like it’s something that we can use to address low-hanging tasks, but not really for advanced work,” he says. “But this paper in particular says that senior developers are more effective, and explains why.” It suggests that “competence is still an important factor that allows you to be augmented with AI,” he adds.

Deadline: January 4, 2026

Call for Proposals for London 2026 is open!

Submit your talk idea

But alongside that, the tech can also be used for gruntwork or translational skills. Fazio says that he often works in code bases across different languages – which can be tricky for him. “Using LLMs has allowed me, for example, to be useful in languages that I wouldn’t be able to use before,” he explains.

Oskar Schulz, president of Cursor, said in a blog post that the company was “encouraged by these early findings, and we’d like to continue studying Cursor’s effects on productivity.” But he also said he was cautious about declaring that this is the full or only impact that AI coding support tools can have. “There isn’t yet a single definitive metric for measuring the economic impact of AI on software engineering,” he wrote. “Like with any new technology, realizing AI’s full value will take time.”