The most interesting AI story this week is not another leaderboard jump. It is a usage pattern: people are starting to treat agents less like chatbots and more like parallel coworkers.
A new arXiv paper, "The Shift to Agentic AI: Evidence from Codex," analyzes usage data from OpenAI's Codex tool and offers one of the clearer public looks yet at how agentic AI is actually being adopted. The authors compare external personal users, external organizational users, and workers inside OpenAI through a privacy-protecting pipeline. Their headline finding is blunt: active Codex users grew more than fivefold in the first half of 2026, and the fastest growth came from people outside the original developer-heavy audience.
That matters because the AI market has spent the last year promising agents that can plan, execute, revise, and coordinate work. Most of that promise has been argued through demos. This paper is more useful because it looks at behavior. Inside OpenAI, the authors say Codex use is nearly universal and has largely replaced business use of ChatGPT. Outside OpenAI, organizational adoption is still lower and uneven, but the same shift toward agentic tooling is visible.
The practical signal is not simply that more people are trying Codex. It is that usage is getting more sophisticated. The paper reports that more than 10% of users manage three or more concurrent Codex agents at some point each week. It also says 26.6% use skills, the reusable instruction bundles that let teams encode workflows instead of re-explaining the same process every time. That is the difference between a clever assistant and an operating pattern.
There is another sharp detail: request complexity is rising. Since the start of 2026, the share of individual Codex users submitting at least one task estimated to take an experienced human more than eight hours has increased nearly tenfold. In plain English, people are becoming more willing to hand agents larger chunks of work. That does not mean the agents finish every job perfectly. It does mean trust, tolerance, and workflow design are changing quickly.
For Daily AI Paper readers, the business implication is straightforward. The next productivity gap may not come from who has access to the best model. It may come from who has redesigned work around multiple agents, reusable instructions, review loops, and clear ownership. A team that uses an agent as a fancy autocomplete will get incremental gains. A team that learns to dispatch, supervise, and compound agent work can change throughput in a more structural way.
The paper also hints at why the transition will be uneven. Agentic tools reward people and organizations that can specify outcomes, break down work, verify results, and tolerate iteration. Those are management and process skills as much as technical ones. Companies that buy seats without changing workflows may see scattered wins. Companies that build repeatable agent practices could create a real operating advantage.
There is a caution here, too. Usage data from OpenAI and Codex is not the entire economy. It is a view into a frontier-tool user base, with OpenAI employees representing an especially unusual environment. The results should not be read as proof that every role is ready to be reorganized around agents tomorrow. But they are strong evidence that the center of gravity is shifting from chat to delegation.
That is why this paper is worth attention. It moves the conversation away from abstract claims about agents and toward the harder question: what happens when knowledge workers stop asking AI for answers and start assigning it work?