The shift from chatbots to agents, measured from the inside
OpenAI published a year of internal Codex telemetry showing the move from chatbots to agents — longer task horizons, non-developer adoption up 137x. Here is what the data shows and where it deserves skepticism.
A data point, not a model
On June 25, 2026, OpenAI published an Economic Research paper on how agents are transforming work, built on a year of internal telemetry from Codex, its coding agent. There is no new model here and no benchmark. What makes it worth reading is that it is one of the first official, quantified looks at a transition many of us have been describing qualitatively: the move from chatbots to agents as the primary way people get work done.
The framing OpenAI uses is the one I would use. Agentic AI changes the unit of knowledge work from a single interaction to a delegated, long-horizon task. A chatbot turn is short and self-contained — you ask, it answers, you are back in the loop. An agent runs on its own for minutes or hours, orchestrating tool calls, touching real environments, and iterating toward a result you check at the end rather than supervise throughout.
That sounds like positioning until you look at the numbers behind it.
The numbers that matter
Three findings stood out to me.
First, the work is getting longer. 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% crossed one hour, and 25.6% crossed eight hours. Nearly a quarter of all requests now correspond to more than an hour of human effort. At the top end, OpenAI's 99th-percentile users generate more than 60 hours of agent turns per day — only possible by running many agents in parallel.
Second, the growth is steepest among non-developers. Since August 2025, non-developer use rose 137x among individual users and 189x among organizational users. Inside OpenAI, Codex became the primary tool for Legal, Finance, and Recruiting around April 2026 — later than Engineering, but the crossover was faster. Research usage in June 2026 was 56x its November 2025 level.
Third, the tool dissolved job boundaries. Over a quarter of the work that OpenAI's business-function staff did with Codex was engineering or coding — work that previously required a ticket to an engineer. The agent lowered the cost of crossing a task boundary, so people crossed it.
My read
I would treat this as a leading indicator, not a finished picture, and I would hold two things in mind.
It is the cleanest evidence yet that the chatbot-to-agent shift is real and measurable, not a narrative. The shape of the curves — non-technical adoption outpacing technical, work horizons lengthening — matches what I see standing up agentic systems in production. When you give capable people low-friction access to an agent that can run for an hour, they stop asking it for answers and start handing it tasks.
But the conditions are not typical. This is one company measuring its own tool, staffed by unusually capable users with frontier access and near-zero adoption friction. The time-horizon figures come from an LLM-as-judge reading transcripts, so they are directional rather than exact, and the user-level numbers rest on a 0.1% sample — OpenAI says as much in its footnotes. An internal share of 99.8% of output tokens on Codex is not where the median enterprise will be next quarter.
What I would take from it: the bottleneck is shifting. Once the agent can sustain an hour of useful work, the constraint stops being raw model capability and becomes the surrounding system — how you scope a task, verify the output, and run agents in parallel without losing the thread. That is the engineering problem worth investing in now, and this paper is the first official dataset that says so out loud.