AI Coding Is Going From Copilot to Autopilot

A new breed of AI-powered coding tools have emerged—and they’re claiming to be more autonomous versions of earlier assistants like GitHub Copilot, Amazon CodeWhisperer, and Tabnine.

One such new entrant, Devin AI, has been dubbed an “AI software engineer” by its maker, applied AI lab Cognition. According to Cognition, Devin can perform all these tasks unassisted: build a website from scratch and deploy it, find and fix bugs in codebases, and even train and fine-tune its own large language model.

Following its launch, open-source alternatives to Devin have cropped up, including Devika and OpenDevin. Meanwhile makers of established assistants have not been standing still. Researchers at Microsoft, GitHub Copilot’s developer, recently uploaded a paper to the arXiv preprint server introducing AutoDev, which uses autonomous AI agents to generate code and test cases, run tests and check the results, and fix bugs within the test cases.

“It’s exciting to see more versions of AI coding assistants with new capabilities,” says Ben Dechrai, a coder and developer advocate at software company Sonar. “They validate the need for generative AI tools in developers’ workflows.”

Dechrai adds that these coding copilots can help software engineers write code faster, allowing them to focus on more strategic and creative tasks. Another advantage of these programming tools is the ability to create a template for code, notes Saurabh Bagchi, a professor of electrical and computer engineering at Purdue University. Much as with prompt engineering, developers must provide these assistants with “the right kind of software requirements to produce a template, and then a software engineer can fill in the gaps,” he says.

“To develop intuitive systems, you need an iterative process with humans in the loop to provide feedback” —Saurabh Bagchi, Purdue University

These gaps include safety and reliability considerations. Software engineers must look out for security vulnerabilities in AI-generated code, as well as the types of corner cases that could cause it to crash.

“Developers still need to ensure rigorous quality standards are in place when analyzing and reviewing code written with generative AI, just as they would with code developed by a human,” says Dechrai. “AI coding assistants are good at suggesting code, reflecting on the code, and reasoning about its effectiveness, but even then it’s not 100 percent accurate.”

Dechrai cautions that autonomous coders are “still so new that developers are just learning which use cases will be most beneficial.” And they’ll need to be “ironed out in the real world to see how much they’re able to deliver on their promise,” says Bagchi.

AI Coders vs. the Humans

Doom-and-gloom predictions of replacing human software engineers are also bound to follow the emergence of these “AI software engineers,” but that won’t be happening anytime soon. Devin, for instance, resolved only 14 percent of a…