AI agents are reshaping work! OpenAI proves it with real Codex usage data

Hey everyone, it's Shii-chan! Today I want to talk about a study that shows how AI agents are actually changing the way we work — not with vague predictions, but with real usage data. There are a lot of numbers here, so let's dig into them together!

OpenAI News openai.com

What was announced?

This is a research announcement posted on OpenAI News. The paper is titled "The Shift to Agentic AI: Evidence from Codex," and there's a friendly companion write-up called "How agents are transforming work." Both went out on June 25, 2026.

What makes it interesting is that OpenAI didn't do this alone — they teamed up with economists from Columbia Business School, the Wharton School at the University of Pennsylvania, and Duke University. And the whole thing is grounded in real data about how the coding agent Codex actually got used.

Why it matters

We hear "AI agents boost productivity!" all the time, but studies that back it up with large-scale, real usage logs are still rare. This paper tracks the shift from November 2025 to June 2026, showing in numbers how agents went from "looks handy" to "the main character of everyday work."

What changes

The biggest change is that the work people hand to agents keeps getting longer and more complex.

Among individual account users, the share who asked Codex for a task that would take a human over an hour rose from 35.4% in December 2025 to 70.2% in May 2026. And the share who handed over jobs that would take more than eight hours grew from 2.1% to 25.6%. That eight-hour tier is close to ten times larger than it was at the start of the year.

The other big shift is that this isn't just an engineer's tool anymore. Even non-technical teams like Legal and Recruiting now use Codex as their primary AI tool. Because agents lower the cost of crossing job boundaries, people can take on "adjacent work" that used to need a specialist.

Dive Deep

Let's look a bit closer at the numbers.

First, user growth: weekly active Codex users grew more than fivefold between January 1 and June 1, 2026, and the fastest growth came from outside the original core of software developers.

Output-token growth by function (how much people had the AI generate) was striking too: inside OpenAI, the median across all functions was at least 10x higher than in November. Researchers grew more than 50x and legal roles about 13x. Broken down by function, one report puts Research around 56x, Customer Support around 32x, Engineering around 27x, and Legal around 13x.

The paper frames this as a move from "asking for information" to "delegating work." Heavy users run multiple agent tasks in parallel like a portfolio, package their common workflows into skills and plugins, and keep longer-running autonomous agents going.

Its conclusion stuck with me: whether productivity actually rises depends not only on how capable the tool is, but on the organizational habits around it — how you delegate, supervise, and verify.

Wrap-up

• OpenAI News published the paper "The Shift to Agentic AI: Evidence from Codex" and a companion write-up (June 25, 2026)
• Done with economists from Columbia, Wharton, and Duke, analyzing real Codex usage data
• Weekly active users grew more than 5x from January to June 2026, driven mostly by non-developers
• The work people delegate is getting longer and more complex: over-one-hour tasks went from 35.4% to 70.2%, and over-eight-hour tasks from 2.1% to 25.6%
• Codex became the primary tool even in non-technical teams like Legal and Recruiting, making cross-boundary "adjacent work" easier
• The key to productivity is not just tool capability but the organizational craft of delegation, supervision, and verification

If you lead a team thinking about adopting agents, or you're just curious how AI is reshaping work, this one's for you!