The research, titled The Shift to Agentic AI: Evidence from Codex, analyzes usage patterns across three groups— everyday Codex users, companies deploying Codex, and OpenAI staff— highlighting a broad organizational impact.

Traditional metrics like DAU and chat volume are being supplanted by measures such as the complexity of tasks, how long agents run, and the degree of parallelism in workflows.

A key implication is that AI usage patterns are moving toward sustained, multi-agent environments rather than single-task, single-agent interactions.

Users are delegating complex, long-horizon tasks to agents, with near tenfold growth in tasks requiring more than eight hours of human effort since the year began.

A new study by OpenAI and academic researchers examines real-world use of Codex, revealing heavy daily usage by workers and extensive multi-agent workflows across individuals, organizations, and OpenAI employees.

There is growing adoption of parallel agent workflows and predefined skills to automate reusable tasks, with skill usage rising from about 5% in March to roughly 27% in June.

Authors stress value comes from task setting, output verification, and coordinating multiple agents, signaling a shift from individual capability to process design and oversight.

The takeaway is that agentic AI is already transforming work processes; those who redesign workflows to leverage agents will prosper, and adoption is accelerating among frontier users.

Productivity gains are substantial, with OpenAI employee median token output rising at least 10x across functions from late 2025 to mid-2026; lawyers up about 13x and researchers over 50x.

Engineering remains a core use case, but non-developer teams— including business, legal, finance, and operations— are increasingly using Codex for technical tasks once requiring specialized support.

Early findings point to productivity gains, altered work patterns, and potential organizational changes driven by agent-enabled coding and automation.

Median OpenAI employees use AI agents about two and a half hours per day, indicating routine integration of AI assistance into daily work.