OpenAI’s next AI agent is a self-testing software engineer that does what humans won’t

OpenAI is working on a new AI agent that is sure to cause quite a stir if the description given by OpenAI CFO Sarah Friar is anything to go by. Friar recently confirmed that the ChatGPT maker will soon be releasing its third AI agent, “A-SWE”, or Agentic Software Engineer. This agent will not only be able to perform the tasks that a normal software engineer can do, but will also be able to take on other additional tasks such as quality assurance, bug testing, and bug bashing.

In an interaction with Goldman Sachs, Friar said, “The third (AI Agent) that is coming is what we call A-SWE… agentic software engineer. And this is not just augmenting the current software engineers in the workforce, which is kind of what we can do today through Copilot, but instead, it's literally an agentic software engineer that can build an app for you.”

Friar added, “It can take a PR that you would give to any other engineer and go build it. But not only does it build it, 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 so, suddenly, you can force, multiply your software engineering workforce.”

OpenAI launched its first AI agent, Operator, in January, soon followed by Deep Research in February. Both AI offerings are currently only available to ChatGPT's paying customers.

Why you shouldn't freak out right now?

OpenAI has a history of making tall claims about its products, some of which never materialise. Take Deep Research, for example. At the time of its launch, OpenAI said that the new tool will be able to replace a research assistant, a claim that Friar repeated on Friday.

While many of OpenAI’s peers, including xAI and Perplexity, have rolled out similar tools, it’s still unclear how much of a research assistant’s role these AI models can genuinely take over. The reason? They're still prone to hallucinations—confidently generating information that simply isn’t true.

The problem isn't that these large language models are inaccurate because, after all, humans make mistakes, too. What’s more concerning is that these models present false information with an air of absolute confidence, making it harder to separate fact from fabrication. And that hasn’t changed much since ChatGPT first rolled out to the public in late 2022.

So when OpenAI says that its upcoming AI agents will be able to essentially do all of what the current software engineers do and more, take those claims with a pinch of salt.