No individual can thoughtfully review sixty hours of delegated work every day.
OpenAI reported on June 25 that its heaviest internal Codex users regularly generated more than sixty hours of agent-turn runtime per day through parallel work. OpenAI says the task-horizon estimates in the study are directional; the sixty-hour figure is reported as Codex runtime. The management problem is concurrency, not a literal conversion to human labor.
Agent capacity can grow faster than review capacity.
A person who dispatches ten tasks at once still has to inspect the results, resolve conflicts, reject bad work, and absorb the decisions. Old review habits were built around one task arriving after another. Parallel agents turn that workflow into a queue.
Teams should measure acceptance without rework, verification time, queue age, discarded output, and errors that escape review. They should set concurrency according to the review capacity available for that class of work.
More agent hours can increase throughput. They can also create invisible work in progress and diffuse responsibility across a pile of plausible artifacts. Utilization is the wrong goal when nobody can tell which outputs deserve to ship.
Each team needs a concurrency limit tied to the reviewers available for that kind of work. Raise it only when acceptance time and escaped-error rates remain stable.