OpenAI published new economic research on June 25 showing how quickly Codex, its agentic coding and work platform, is shifting from a developer tool into a broader workflow system for delegated work.
The company summary and the accompanying research paper, written by researchers from OpenAI, Columbia Business School, Wharton, and Duke, offer one of the more concrete public looks yet at how AI agents are being used in real workplaces. The headline is not simply that more people are trying Codex. It is that heavy users are giving agents longer tasks, running them in parallel, and applying them outside traditional software engineering.
That makes the report more useful than another broad claim that AI agents will transform work someday. OpenAI’s data suggests the shift is already visible among early adopters, but uneven: Codex is deeply embedded inside OpenAI, meaningfully used by some organizations, and still comparatively niche among individual users.
Codex Is Moving From Chat to Delegated Work
The research frames the difference between chatbots and agents around the unit of work. A chatbot interaction is usually a short exchange: ask a question, get an answer, refine the result. Codex-style agentic work is closer to delegation. A user gives the system a goal, and the agent can inspect files, run commands, make changes, test results, and iterate toward an outcome.
OpenAI says weekly active Codex usage grew more than fivefold in the first half of 2026. Among sampled individual users, 80.6% made at least one Codex request estimated to represent more than 30 minutes of experienced human work by May 2026. About 70.2% made at least one request estimated at more than an hour, and 25.6% made at least one request estimated at more than eight hours.
Those estimates should be read carefully. The paper says task length is model-estimated, not directly observed, and the individual-user figure is based on a random sample of 0.1% of users. Still, the direction is important: agent usage is not only getting more frequent, it is getting more ambitious.
The Internal OpenAI Numbers Are Extreme, but Useful
The most dramatic findings come from OpenAI’s own workforce, which is not a normal adoption environment. Employees have unusually low friction, deep product familiarity, strong internal encouragement, and broad access to the systems being studied. The paper is explicit that internal usage is not representative of an average company today.
Even with that caveat, the internal numbers show what agent-heavy work can look like when the barriers are low. OpenAI reports that Codex now accounts for more than 85% of output tokens for the average OpenAI worker, and 99.8% of weekly output tokens generated across Codex and ChatGPT inside the company. Engineering moved first, but legal, finance, and recruiting reportedly crossed into majority Codex usage around April 2026.
The paper also says usage intensity rose sharply across departments. Median combined output tokens for research workers were 56 times higher in June 2026 than in November 2025. Customer support rose 32 times, engineering 27 times, and legal 13 times.
That does not prove productivity rose by the same amount. Tokens are not output quality, revenue, shipped features, or saved hours. But they do show that some workers are reorganizing their day around agent execution rather than treating AI as a side chat window.
Non-Developers Are the Fastest-Growing Group
Codex still has a strong software center of gravity. The research says software work remains the largest use case, including implementation, debugging, refactoring, validation, configuration, repository work, and documentation. That is expected for a tool that began as a coding agent.
The more interesting shift is who is using it. Since August 2025, OpenAI says non-developer individual users grew 137 times, non-developer organizational users grew 189 times, and non-developer users inside OpenAI grew 12 times. Those figures start from different baselines, but they point to the same trend: agentic tools are spreading beyond the engineering department.
The paper’s examples are practical rather than futuristic. Non-technical users are delegating work such as data transformation, structured analysis, document drafting, workflow automation, debugging, and internal tooling. In OpenAI’s business functions, more than a quarter of Codex output tokens were classified as engineering or coding work, even though the employees were not primarily engineers.
That is the real workplace implication. Agent tools do not only speed up tasks people already know how to perform. They can let workers cross into adjacent technical work that previously required a specialist, provided the organization has enough review, permissions, and verification around the output.
Parallel Agents Change the Management Problem
The research also shows that intensive users are not just asking harder questions. They are learning to manage multiple agent runs as a workflow. More than 10% of users manage three or more concurrent Codex agents at some point each week, according to the paper, and 26.6% use skills, which package reusable instructions for complex workflows.
OpenAI’s company post says the heaviest internal users at the 99th percentile were regularly generating more than 60 hours of Codex agent turns per day by June 2026, spread across multiple parallel agents. That figure is less about a single person working superhuman hours and more about the interface changing: the user becomes a coordinator, reviewer, and task designer.
That shift creates new bottlenecks. Teams need clear task specifications, access boundaries, test environments, audit trails, data controls, and human review paths. A person who can run ten agents at once can also create ten streams of flawed work if the organization has weak validation habits.
Why This Matters for Businesses
Axios, which reported on the research Thursday, framed the change as AI moving from chat and search toward delegated work. That is the right lens for companies evaluating agent tools. The question is no longer only whether employees are allowed to use AI. It is what kinds of work they are allowed to delegate, what systems an agent can touch, and who is accountable for the result.
The OpenAI paper also makes clear that adoption depends on organizational context. Individual users still rely mostly on conversational AI, and even among organizations, agent adoption is uneven. Access to files, internal systems, training, review processes, permissions, and management expectations all affect whether an agent becomes useful infrastructure or an occasional experiment.
For technology leaders, the useful takeaway is practical: agent readiness is becoming a workflow-design problem. Companies that want value from tools like Codex need more than licenses. They need repeatable tasks, clean repositories or data sources, safe execution environments, permission models, evaluation routines, and a culture where workers know how to supervise delegated work instead of blindly accepting it.
The agent era may not arrive evenly, and OpenAI’s own workplace is a best-case adoption environment. But the new Codex data gives the trend sharper edges. The early frontier is not a chatbot that answers better. It is a work system where humans increasingly assign, monitor, and verify streams of machine-executed tasks.
