On June 25, OpenAI published a post on how agents are transforming work, drawing on a paper by Drew Johnston and five coauthors from OpenAI, Columbia, Wharton and Duke. One figure concerned OpenAI’s own staff: by June 2026, internal users at the 99th percentile routinely logged more than 60 hours a day of Codex agent turns, spread across agents running in parallel.
The post presents this growth as evidence of what agents can do for work. It says agents move knowledge work from one-off exchanges toward delegated tasks that run for hours, and it describes the result as “an exponentially growing amount of agentic labor.” As a description of use, that holds up: more work is being handed to agents, and agent time is a real input. My disagreement is narrower. Growth in runtime and tokens cannot establish that output, productivity or economic value grew with it, and the post’s framing around economic potential invites that leap.
The paper defines the measure in a footnote. It adds up the time of each request and response inside an agent turn, including gaps of up to 30 minutes while tools run or the model works, and drops longer gaps, which mostly mean the agent is waiting for the user. Batch-script turns are excluded, parallel turns are summed, and accounts inactive on a given day are left out, so one person’s daily total can exceed 24 hours. Among OpenAI employees active on a given day, the 99th percentile was about 71 hours in the paper’s data, and the median active employee on June 11 had turns running for 2.5 hours.
The number therefore measures how much agent activity one person starts and keeps running, including compiling, running tests, waiting on tools and retrying. The paper reads the pattern as a shift toward delegating and supervising, which is a reasonable interpretation, though runtime alone cannot show how people spend their working hours. It also does not mean sixty hours of output waiting for a person to read, since some of the runtime is the agent checking its own work against tests.
On what the activity produces, the paper’s authors argue that agentic AI pays off only where organizations rebuild workflows and review processes around handing work off and checking it. They cite a working paper by Mert Demirer, Leon Musolff and Liyuan Yang, revised in September 2026, which matches more than 500,000 GitHub developers with their AI usage. Its matched event-study estimates put the cumulative effect of autonomous coding agents at 240 percent for coding activity, measured as commits, but 80 percent for the number of projects and 30 percent for releases. In four large software marketplaces, many more new apps appeared without any rise in overall use. The authors read the falloff as people further down the line limiting how much extra code becomes shipped software. Runtime sits at the top of that chain, where the measured effects are largest.
Whether a person has to look depends on the work. Automated checks can carry the load when the requirement is written in a form a machine can test (a test suite, a type checker, a reconciliation that must tie to the ledger), when the agent cannot edit the test to pass it, and when a miss can be reverted cheaply. A dependency upgrade in a well-tested codebase fits: a person reads a summary and the results. A person has to read the output when the requirement lives in someone’s judgment, such as whether a contract summary caught the clause that matters, when the output acts on the world, or when an error would surface only much later. For the contract summary, checking means reading the contract, which can cost nearly as much as writing the summary.
The paper explains that Codex was first built for software development, a domain it describes as “comparatively easy to verify.” The post’s own breakdown shows how far use has spread beyond it: for product, marketing and operations staff at OpenAI, knowledge work accounted for 51 percent of Codex output tokens and engineering or coding for 25 percent. My inference, which the paper does not test, is that as delegation moves into that kind of work, an hour of runtime will need more human review than it does in code.
Neither the post nor the paper reports how much agent output was accepted, reworked or discarded, or how many errors got past review; the paper suggests that future analyses track “production output.” Until acceptance and error rates are published per hour of runtime and by type of work, nobody outside OpenAI can tell whether its heaviest users are its most productive or its most active. Independent measurement is getting harder as well: METR reported in February that its follow-up developer study could no longer give a reliable estimate, partly because time per task stops meaning much when developers run several agents at once. If OpenAI published acceptance rates per agent-hour by department, and its legal and finance teams kept most agent output after light review, I would conclude that verification is not yet the constraint I think it is, at least for the work those teams hand off.
Revised October 11, 2026. The version published June 27 argued that no one could review sixty hours of delegated work a day, treating runtime as reviewable output. This version uses the paper’s definition of the measure and replies to the argument the post makes.