Inside OpenAI, AI Agents Have Replaced the Chatbot

In Brief

OpenAI released new research, co-authored with economists from Columbia, Wharton (at the University of Pennsylvania), and Duke, that uses real usage data from its Codex agent to show how agentic AI is reshaping work. 

Active Codex users grew more than 5x in the first half of 2026. The fastest growth came from outside software engineering. Inside OpenAI, a single agent now drives 99.8% of weekly output tokens. The unit of knowledge work is shifting from quick chats to long, delegated tasks.

What Happened

OpenAI just published new research, co-authored with economists from Columbia, University of Pennsylvania's Wharton School, and Duke, that uses real usage data from its Codex agent to show how agentic AI is changing the way people actually work. Codex is OpenAI's AI agent that can do work for you, not just chat. Agentic AI means AI that takes multiple steps on its own to complete a task instead of just answering a single question. The findings appear in a research paper, "The Shift to Agentic AI: Evidence from Codex,"and a companion post, "How agents are transforming work."

The headline finding is that agents are spreading fast and well beyond engineers. Active Codex users grew more than fivefold in the first half of 2026, with the most rapid growth coming from outside the original audience of software developers. People are also handing agents bigger jobs. The share of individual users who gave Codex at least one task that would take an experienced human more than eight hours grew nearly tenfold since the start of the year. More than 10% of users now run three or more Codex agents at the same time in a given week, and about a quarter use saved "skills," which are reusable instructions for complex workflows. Usage is not capped inside OpenAI.

The research offers a glimpse of where this is headed. Codex now drives more than 99% of employee work output and has largely replaced ChatGPT. This past month, the median employee in a legal role at OpenAI produced 13 times more across the company's AI tools than in November. The median researcher produced more than 50 times as much.

On Episode 222 of The Artificial Intelligence Show, SmarterX founder and CEO Paul Roetzer looks at what this shift means.

The Key Numbers

99.8% - Share of OpenAI's weekly output tokens now from Codex

5x - Growth in active Codex users in the first half of 2026

26% - Individual users who gave Codex an 8+ hour human-work task

50x - Output increase for OpenAI's median researcher since November

81% - Users who made a 30+ minute human-work request by May 2026

Why the Unit of Knowledge Work Is Changing

The shift is from answering to acting. That is the core distinction the research draws. As OpenAI put it: "Agentic AI changes the unit of knowledge work from single interactions to delegated long-horizon tasks." Long-horizon tasks are jobs that take many steps over a long stretch of time. OpenAI described it this way: "Chatbot interactions are often short and self-contained. Agents can operate independently for minutes or hours while orchestrating tool calls, interacting with environments, and iterating towards solutions. As a result, agents are quickly becoming the most powerful AI tool for work."

The data inside the lab tells the story. By May 2026, the research found, 81% of sampled individual users made at least one Codex request estimated to exceed 30 minutes of human work, 70% made one estimated to exceed an hour, and 26% made at least one estimated to exceed eight hours. Engineering moved to Codex first, but legal, finance, and recruiting all crossed over to it as their primary tool around April 2026. Over the past six months, usage rose 32 times in customer support, 27 times in engineering, and 13 times in legal.

Roetzer expects companies to follow the labs. "This research to me is invaluable, whether it's coming from here. Anthropic has put out some great research just on their own internal usage, because all of us are trying to figure out what the future of work looks like and how it's gonna reshape our organizations, our tech stack, our org charts," says Roetzer. "And the best chance to see around the corner is to look inside of the labs who are actually using all this technology and have, in this case, infinite token use basically." Tokens are the chunks of text AI is measured and billed in, and the labs are not rationing them the way most companies must.

The mental model is simple and unsettling. Roetzer offered a way to picture it. The average person works roughly eight hours a day. Imagine swapping that day for a handful of prompts, where each prompt produces what used to take an hour. Then scale it up.

"We're entering a phase where agentic capabilities are going to be made accessible that can do hours or days worth of work with simple prompt strings that could be completed in minutes. And that's just fundamentally a really difficult thing to comprehend."

— Paul Roetzer, founder and CEO of SmarterX, Episode 222 of The Artificial Intelligence Show

SmarterX Take

There is a stubborn misconception buried in this story, and it is worth clearing up. Codex and Claude Code are sold as coding tools, but that is not the only way they are used. They function as general-purpose agentic harnesses wrapped around the frontier models, and people run them every day to do ordinary knowledge work, not to write software. That is exactly why legal, finance, and recruiting teams inside OpenAI adopted Codex as their primary tool. You do not need technical knowledge to assign an agent a long, multi-step task and walk away.

This is why agentic work is becoming SmarterX's core research focus for the second half of the year, and a focus of its AI council. "One of the most consequential things that will happen in business is the autonomy and reliability of agents, and then finding out how to get affordable access to do what they make possible efficiently," says Roetzer. The capability is arriving faster than most organizations are ready for, and the easiest way to understand it is to start using these tools yourself.

What to Watch

Adoption is moving fastest where no one expected it. The original audience for agents was developers, but the steepest growth is now coming from non-technical roles. As more business professionals discover that an agent can quietly handle a day's worth of analysis, drafting, or research, expect the line between "coding tools" and "work tools" to disappear entirely.

Affordable access is the next bottleneck. The labs run on essentially uncapped token budgets, which allows them to run so many AI agents. The question is how everyone else gets affordable, reliable access to the same long-horizon agentic work without their bill spiraling out of control. The companies that solve that first will get the productivity gains first.

Professionals Want to Learn about AI Agents

When asked what AI topics they want training on, 51% of professionals named using AI agents, second only to integrating AI tools into existing workflows at 58%, according to the 2026 State of AI for Business Report. The demand is there. The skills, for most teams, are not yet.

The report is built on more than 2,100 responses across roles, functions, and industries, and it maps exactly where organizations sit on the path from understanding AI to scaling it. If the research from OpenAI is a preview of how work changes when agents take over long-horizon tasks, this report is the benchmark for figuring out how ready your team actually is. Read the full report →