Generalized LLMs were already looking poised to change how coding works, but OpenAI is looking to create a specialized agent that’s even better at coding tasks.
OpenAI CFO Sarah Friar has offered a glimpse into the future of software development, hinting at a groundbreaking OpenAI project codenamed “Agentic Software Engineer” – or “A-SWE”. Friar’s comments suggest a significant leap beyond current AI coding assistants like GitHub Copilot, envisioning an autonomous AI capable of handling the entire software development lifecycle.
“The third (product) that is coming,” Friar said in a chat today, “is what we call A-SWE. We are not the best marketers, by the way, you might have noticed, but [it’s an] Agentic Software Engineer,” she laughed at the A-SWE name.
Friar said that A-SWE wasn’t the same as how current LLMs were improving coding performance of human engineers. “This is not just augmenting the current software engineers in your workforce, which is kind of what we can do today through Copilot,” she said. Instead, Friar painted a picture of a fully autonomous AI developer: “It’s literally an agentic software engineer that can build an app for you. It can take a PR [pull request] that you would give to any other engineer and go build it.”
The most compelling aspect of A-SWE, however, lies in its ability to handle the less glamorous, but essential tasks of software development. “But not only does it build it,” Friar emphasizes, “it does all the things that software engineers hate to do. It does its own QA, its own quality assurance, its own bug testing and bug bashing, and it does documentation—things you can never get software engineers to do.” The result, she suggests, is a dramatic increase in development capacity: “So suddenly you can force multiply your software engineering workforce.”
There have been a few attempts at making a AI Software engineer. Most prominently, Peter Thiel-backed Devin had been launched last year, and was able to not only to respond to Slack messages from its human managers, but also write and publish code. Reactions to Devin had been mixed, and people had found its price of $500 a month to be somewhat steep. Also, Devin wasn’t nearly as autonomous as one would’ve expected, and frequently broke down or created buggy code.
But OpenAI entering the AI coding agent space could shake things up a bit. OpenAI not only has some of the best AI models out there, but has also received plenty of feedback on the code its LLMs have created over the last couple of years. That, along with its massive distribution — it now has over 500 million weekly active users — could really enable it to push a capable AI software agent into the mainstream.
And if OpenAI’s A-SWE can really understand requirements, write code, do QA, and even write documentation as its management suggests, it could end up upturning the coding industry. OpenAI CEO Sam Altman has previously said that the world’s best programmer will be an AI by the end of 2025. And if this AI can be deployed and perform tasks like human engineers, it could massively impact the software engineering job market in the coming years.