OpenAI published “The Shift to Agentic AI: Evidence from Codex,” a research paper co-authored with economists from Wharton (University of Pennsylvania), Columbia Business School, and Duke University. The study draws on real-world usage data from millions of Codex interactions across three groups: individual users, organizational users, and OpenAI’s own employees.
The core finding is blunt. At the frontier of adoption, agents have nearly completely replaced chatbots as the primary interface for knowledge work. Among OpenAI employees, 99.8% of output tokens now flow through Codex rather than traditional ChatGPT-style chat. For organizational users the figure is 63.3%. Individual users sit at 16.5%, according to OpenAI’s research blog.
Task Complexity Is Surging
The paper documents a steep rise in the difficulty of work being delegated to agents. Among individual users, the share submitting requests estimated to require more than eight hours of human effort grew from 2.1% to 25.6% since January 2026. By May, 80.6% of sampled individual users had made at least one request estimated to exceed 30 minutes of human work, and 70.2% made one exceeding an hour.
Users front-load the most complex parts of a task at the beginning of a session, then iterate with refinements, a pattern consistent with treating agents as delegated workers rather than interactive advisors.
Non-Developer Adoption Outpaces Engineering
Engineering remains the primary use case, but the fastest growth is elsewhere. Non-developer adoption rose 137x for individual users and 189x for organizational users since August 2025, according to Quasa’s analysis of the paper. Legal, Finance, and Recruiting teams at OpenAI crossed into Codex as their primary AI tool around April 2026.
Among organizational engineers specifically, 26.8% of tokens now flow through Codex, five times higher than in January. For active users in that group, the figure reaches 88.3%.
Parallel Agent Workflows Emerge at Scale
The paper tracks the emergence of sophisticated multi-agent operations. At OpenAI, 28.6% of users managed five or more agents simultaneously during a recent week. Users at the 99th percentile generated more than 60 hours of Codex agent runtime per day, distributed across parallel agents.
The use of “skills,” pre-defined instructions for complex reusable workflows, grew from 5.4% of users in March to 26.6% in June. At OpenAI internally, 96.2% of employees invoke at least one skill. Median token output per OpenAI employee increased at least 10x from November 2025 to June 2026. Lawyers saw 13x growth; researchers more than 50x.
The Electrification Analogy
The authors draw a parallel to factory electrification: productivity gains came not from swapping steam engines for electric motors, but from redesigning entire factory layouts around the new technology. With agentic AI, the paper argues, “rebuilding costs are negligible” since restructuring workflows requires little more than time and imagination, suggesting the transition will outpace previous industrial shifts.
New Metrics Needed
The paper argues that traditional engagement metrics like daily active users and message counts are becoming obsolete for measuring AI’s economic impact. The relevant measures are now task complexity, agent runtime duration, and degree of parallelism. Human value is shifting from “being able to do the work” to “setting the right tasks, verifying outputs, and coordinating multiple agents.”
For agent infrastructure builders, the data validates a thesis many already operated on: the market is moving from conversational AI toward autonomous delegation at a pace that even the companies building these tools find surprising.