Inside OpenAI, This Productivity Hack Is Giving Workers Their Own Chief of Staff. You Can Use It Too

Inside OpenAI, the company behind ChatGPT, employees both technical and non-technical are using Codex, the company’s agentic coding app, to handle an increasing amount of work. Codex is OpenAI’s label for its AI-coding platform, which has been accessible through the cloud for over a year. It experienced a huge surge in growth, however, following the release of a dedicated Codex desktop app for Mac and Windows PCs. Unlike the cloud-based versions of Codex, the desktop app is capable of connecting to a computer’s local filesystem, editing existing files, and creating whole new files. In effect, this means that by using the Codex app, users can direct AI agents to do any work that requires a computer. Within OpenAI, non-technical employees are using it in ambitious ways. “The investor relations team is like three people,” says Alex Embiricos, the product manager responsible for Codex, and has been using Codex to monitor the influx of cash from the company’s recent $122 billion fundraise. During a recent all-hands meeting, Embiricos recalls, CFO Sarah Friar told staff that a member of her team had about half an hour to kill, and in that time “vibe coded a dashboard that just showed the incoming expected transfers lighting up, and sent it to Sarah over the weekend.” Two members of the Codex team spend much of their time hosting office hours and answering questions from non-engineering teams at OpenAI. By monitoring if usage rates spike after one of these sessions or hackathons, the Codex teams can understand more about what use cases certain divisions of the company are discovering for themselves. Laura Peng, one of the members of the Codex product operations team, says that many employees are getting utility out of simple dashboards that connect to their email and work communications service like Teams, Slack, and Google Meet, along with data sources like Excel and Sheets. Employees are monitoring emails, keeping track of deadlines, and summarizing and responding to Slack messages. “I feel like I almost have my own chief of staff, in a way,” says Peng, “just making sure that I’m staying on top of my job.” As OpenAI’s ever-growing number of employees get up to speed with Codex, Peng says, “the floor has just [been] raised for everyone, where people’s level of curiosity about a tool they previously would have been really intimidated by has just grown.” Personally, Peng is currently planning a trip to Korea, and used Codex to create an interactive interface where she could “kind of click on the different cities that we were visiting and see how far everything was from each other.” One popular method for learning how to get the most out of Codex, says Peng, is developing small games. Maybe you start by making a classic Snake game, she suggests, and then ask Codex to make the snake neon pink, and then turn the targets that the snake eats into digital strawberries. This kind of rapid iteration is fantastic for teaching people about the open-ended problems that Codex can solve. Katy Shi, a research lead on Codex, also creates games with Codex, but usually as a means of benchmarking the agent when new updates or models are released. One popular challenge is to see how Codex reacts when being tasked to make a first-person version of Tetris from the perspective of the piece. Part of Shi’s work involves crafting the personality of Codex, and finding a way to thread the line of being helpful without becoming too trigger-happy or overeager in its actions. To be sure, OpenAI isn’t the only company that is bringing the benefits of agentic coding to non-engineering work. In January, the company’s chief rival Anthropic released Claude Cowork, a feature on the company’s Claude desktop app that takes some of the abilities of its popular Claude Code product, and packages them in a more beginner-friendly user interface. Instead of creating two separate products for engineering and knowledge work, OpenAI has elected to make Codex adept at handling both types of tasks. Recently, OpenAI has begun adding some aspects of Codex back into ChatGPT, such as Workspace Agents in ChatGPT, which take the form of little animated characters that workers can assign to handle specific tasks, and then share among their organization. Eventually, the plan is for Codex and ChatGPT to merge into a single “super-app” that handles everything for you in a simple interface. One-off features like Workspace Agents are just a taste of what’s to come in the future. For now, Embiricos is laser-focused on making the Codex desktop app as good as it can possibly be, and those efforts are paying off. Since the February 2 launch of the Codex desktop app, OpenAI says, the app has grown to over 4 million weekly active users. At some point, Embiricos says, work will shift over to making the mobile version of Codex as useful as the desktop version, and then the need for desktops and laptops will eventually just go away. “I don’t even want to have to open my computer,” says Embiricos. It won’t be long before everyone in the world has a true “personal assistant on their phone.” BY BEN SHERRY @BENLUCASSHERRY