AI is helping people write code, but more and more of the code at these AI companies is now being written by AI itself.
This revelation comes from Alexander Embiricos, a key figure at OpenAI who joined the company after it acquired his developer collaboration startup, Multi. In a recent discussion, Embiricos, who is part of OpenAI’s Codex team, painted a picture of a rapidly evolving software development landscape, one where human programmers are increasingly becoming orchestrators of AI agents that do the bulk of the coding.
Embiricos foresees a fundamental shift in how code is created within the next two years. “In the next two years, coding will look completely different,” he stated. He elaborated on the current state of AI in programming, noting the prevalence of AI tools that act as direct assistants to developers. “I think right now most of the tools that people find the most value from are tools that work really closely with you, like in your development environment. You know, basically pairing.”
However, this collaborative model is on the verge of being superseded by a more autonomous one. “The shift that we’re gonna see… is that actually the majority of code will be written by agents,” Embiricos explained. He emphasized a crucial distinction in how these agents will operate: “And those agents won’t be working in your environment where you can do one thing at a time, but they’ll be working in their own environments and they won’t just be triggered by you, thinking of specific tasks, but they’ll be connected into the tools you use doing work there.”
This move towards autonomous AI agents is not a distant future for OpenAI; it’s already happening. Embiricos revealed the extent to which these agents are contributing to their codebase. “Even already at OpenAI, we’re seeing like, much more code is merged by agents,” he said. He also pointed to a fascinating new dynamic in their development process: “Even more code is generated by agents as folks are kicking off tasks four times to choose their favorite implementation.”
This proliferation of AI-generated code introduces new challenges. The sheer volume and the novel methods of its creation are forcing a rethink of established software development practices. “And so it’s like not a hundred percent clear how we should even manage all this code that is being written,” Embiricos admitted. He also highlighted the looming question of how to ensure the quality and reliability of code written by autonomous agents. “I think we’re gonna have to figure a lot out, a lot about code review… personally, like I don’t exactly know how that’s gonna work.”
The implications of this shift are profound. The traditional role of a software developer is set to transform from a hands-on creator of code to a high-level strategist and quality controller — this is something a researcher from Anthropic had recently said. Instead of writing line-by-line instructions, developers will be defining problems, setting goals for AI agents, and then evaluating and integrating the solutions they generate. This trend is echoed by the broader industry’s move towards what is being termed “agentic AI,” where autonomous systems are capable of planning, executing tasks, and learning with minimal human intervention.
We are already seeing the early stages of this transformation with the increasing adoption of advanced AI coding assistants. As AI models become more capable, their role is naturally expanding from providing suggestions to undertaking entire development tasks. This is evident in the development of sophisticated AI agents by companies across the tech landscape, all aiming to accelerate development cycles and unlock new levels of productivity. The challenge, as Embiricos candidly pointed out, will be in building the new infrastructure and processes required to manage this new era of software creation.