Codex Data Shows Agentic AI Moving From Chat to Delegated Work
OpenAI's new economic research paper uses Codex usage data to show how agentic AI is shifting work from short chatbot exchanges toward longer, delegated, tool-using workflows.
For years, the word "agent" was mostly a promise. A new OpenAI economic research paper gives that promise something more useful: usage data.
On June 25, 2026, OpenAI published "The Shift to Agentic AI: Evidence from Codex", a paper written with researchers affiliated with Columbia Business School, the Wharton School at the University of Pennsylvania, and Duke University's Fuqua School of Business.1 The paper studies Codex usage across three groups: individual users, organizational users, and OpenAI workers.2
The important distinction is simple. Chatbots usually answer. Agents do work. In the paper's framing, agentic tools can inspect files, use external tools, execute commands, create or modify artifacts, and operate on a user's behalf across multi-step tasks.2 That changes the unit of work from a single prompt-response exchange into a delegated workflow.
What the data says
The headline finding is that Codex usage grew quickly in early 2026, but unevenly. The paper says active Codex users grew more than fivefold in the first half of 2026.2 OpenAI's own workplace is the most extreme case: by June 11, 2026, Codex accounted for 99.8% of output tokens generated by OpenAI workers across Codex and ChatGPT.2
That internal number is striking, but it should not be read as a normal-company benchmark. The paper explicitly notes that OpenAI is an unusually favorable environment: workers are close to the technology, marginal usage is cheap internally, organizational support is high, and informal knowledge sharing is common.2 In other words, OpenAI is useful as a preview of what low-friction agent adoption can look like, not as proof that every company will follow the same curve immediately.
The broader trend is still meaningful. OpenAI reports that Codex became the primary AI tool for every department inside the company, including non-technical departments such as Legal, Finance, and Recruiting.1 Non-developer usage grew especially fast: since August 2025, OpenAI says non-developer users increased 137x for individual users and 189x for organizational users.1
The nature of the work also changed. By May 2026, 80.6% of sampled individual users had made at least one Codex request estimated to exceed 30 minutes of human work, 70.2% had made one estimated to exceed one hour, and 25.6% had made one estimated to exceed eight hours.1 OpenAI cautions that these task-duration thresholds are model-estimated and based on a 0.1% random sample of individual users who allowed their queries to be used for training.1
Why this matters for software teams
Codex started as a coding tool, and software work remains central. But the paper describes usage beyond code generation: debugging, refactoring, validating changes, configuring applications, drafting documents, analyzing data, and coordinating communication.2 That is the real shift. The agent is no longer just a smarter autocomplete surface. It becomes a workbench where a human delegates, monitors, reviews, and coordinates execution.
For developers, this changes the practical skill set. Prompting still matters, but the bigger leverage is in making work delegable: clean repositories, runnable tests, explicit acceptance criteria, reliable scripts, documented workflows, and safe access boundaries. Agentic AI rewards teams that already know how to package work clearly.
It also changes the bottleneck. If a developer can run several agents in parallel, the scarce resource becomes review quality, system understanding, and judgment about what should be automated at all. The paper notes that more than 10% of users manage three or more concurrent Codex agents at some point each week, and 26.6% use skills for reusable workflow instructions.2 That points toward a future where software teams manage parallel streams of delegated work instead of only writing code line by line.
The sober reading is not "agents have replaced workers." The evidence is narrower and more interesting: in environments where agentic tools are available, trusted, and integrated into real workflows, users start assigning them longer, more complex, and more cross-functional tasks. That is a different adoption pattern from classic chatbot use.
For organizations, the takeaway is practical. Buying access to an agent is not enough. The productivity gain depends on whether the surrounding workflow is ready for delegation, verification, and iteration. The companies that benefit most will probably be the ones that treat agents less like magic employees and more like powerful, inspectable execution systems.