[ad_1]
OpenAI has shipped new products at a relentless clip in the second half of 2025. Not only has the company released several new AI models, but also new features within ChatGPT, an AI-powered web browser, and Sora 2, its AI video-generation and social media platform.
The secret behind OpenAI’s lightning-fast shipping cadence can be summed up in a single word: Codex.
Codex is OpenAI’s family of agentic coding tools. These tools allow AI agents to write, edit, and run code at a speed and scale that’s simply impossible for humans to match. Codex product lead Alex Embiricos says that his end goal is for the product to recreate the feeling of working closely with a highly-skilled software engineer—one that can work for hours, even days on end.
Codex has been well received by the coding community; Embiricos says that daily usage of the system has jumped by 10x since August, when the company released GPT-5, a model that was considered a major step up for OpenAI in the coding department. Software engineers and nontechnical “vibe coders” can access Codex as an extension in dedicated coding apps like VSCode and Cursor, through desktop-based coding terminals, and on the web via the cloud.
To be sure, Codex isn’t the only agentic, AI-powered coding tool out there. Anthropic’s Claude Code tool was released in February and has been widely adopted by professional developers; Google’s Gemini can be used to develop apps; and startups like Base44 (now owned by Wix), Lovable and Replit have developed user-friendly platforms for agentic coding.
Internally, Embiricos says around 92 percent of OpenAI’s engineers are making heavy use of Codex. The coding agent plays a major part in all of OpenAI’s product launches, and has transformed OpenAI’s product review process. This fall, as OpenAI was preparing to launch video-generation app Sora, Embiricos says that some of the product’s engineers held a debate as to whether they should build a direct messaging feature into it. “Codex basically built the feature in the background while they were debating it,” says Embiricos.
Codex wrote roughly 85 percent of the Android version of the Sora app, says OpenAI engineer Patrick Hum. Not long after Hum was hired, he and three fellow engineers were tasked with converting the Sora app from iOS to Android within a month. To hit the ambitious deadline, Hum says each engineer ran multiple Codex instances simultaneously, essentially turning their team of four into a team of 16.
Hum says those 16 virtual engineers were arguably more effective than a team of 16 human engineers, because instead of aligning all those people around a shared vision, only the four humans needed to be on the same page. This allowed the team to operate at a greater velocity without needing to stop and regroup as often.
Initially, Hum’s team gave Codex access to the Sora iOS app’s code base, and challenged the agent to create the Android version all by itself. After working continuously for around 12 hours, Hum says, Codex delivered a version of the app that “certainly wasn’t anything that we could show anybody.” Engineers across OpenAI, including Hum, refer to Codex as being roughly equivalent to “a senior software engineer that just got hired. It has absolutely no concept of what our best practices are; it doesn’t know about our product plans.”
After the initial, fully-automated approach failed, Hum’s team spent the project’s first week writing code by hand, fleshing out the app’s internal architecture and defining best practices in text files. Hum says this process gives Codex a “context-rich environment” in which the agent can operate with a more complete picture of what it is building.
Working in lock-step with Codex, Hum’s team delivered a functioning version of the app to OpenAI employees in 18 days, and officially launched on the Android app store 10 days later.
John Nastos, an OpenAI engineer working on ChatGPT Atlas, the company’s AI-powered internet browser, estimates that he uses Codex to write 99 percent of his code. Nastos says that Codex was instrumental in developing Atlas’ “agent mode” feature, a complex system that enables an AI agent to take control of the browser, operating its own cursor and taking actions. According to Nastos, the engineering team felt strongly that agent mode needed a logged-out version in which the agent doesn’t have access to users’ cookies and credentials, akin to an incognito mode in a normal browser.
“The feature was complicated,” recalls Nastos. To make it work, the engineers would need to work on three codebases at once. The app’s native client uses Swift, a coding language used to develop applications for Apple hardware, while its web code is written with JavaScript and its backend code is written with Python.
If Nastos was the only engineer working on the agent mode feature, he says, “I’d have to make a lot of decisions about what order I’m going to touch these things in.” But with the advent of tools like Codex, Nastos was able to generate a unified plan for implementing the feature, distribute that plan to three different agents, and have each one work in its own code base.
Once the code was written, the Codex agents were able to run the code, see how each piece worked together, and make adjustments. Nastos says the feature was almost entirely written and tested by Codex, and would’ve taken an estimated four times as long to do by hand.
As Codex continues to improve and gain popularity, Embiricos anticipates the agentic system will become more intuitive to people who don’t possess any software engineering skills. For non-coders who don’t want to wait, Embiricos suggests asking Codex (one good way for a non-expert to do that would be to use the Codex extension in a coding app like Cursor) for help in instances where you’d otherwise contact your engineering team, and using the system to develop simple side projects that appeal to your personal interests.
“Just try to make something good,” adds Embiricos. “You don’t need to overthink it.”
The final deadline for the 2026 Inc. Regionals Awards is Friday, December 12, at 11:59 p.m. PT. Apply now.
[ad_2]
Ben Sherry
Source link