An agent can spend an hour writing code while another spends that same hour testing it. Two machine-hours have elapsed. A person has lived through one. Neither count tells us whether useful software reached a customer.
Anthropic's October 16 employment report makes that distinction unusually concrete. By August, employee-directed agents logged six times estimated human working hours. The calculation pools subagents, inference and active tool use, excluding waits longer than 30 minutes. Human hours are headcount multiplied by an assumed 40-hour week, not measured time completing equivalent tasks.
Six-to-one measures runtime, not six times the output or six people replaced.
More running, more waiting
The same report finds that interactive sessions spend roughly 80% of their total elapsed time waiting for human input. That is a different denominator from the runtime calculation; the figures cannot be subtracted from each other.
Employees increasingly build systems to organize agents and their requests. Producing material has created more coordination work.
This is the AI vendor studying its own fast-growing company. Tool adoption was not randomized. The observations do not establish how much productivity AI caused, or how other employers will fare.
A code change is not a release
A separate working paper by Mert Demirer, Leon Musolff and Liyuan Yang follows the distinction further downstream. Its September 2026 revision combines AI-use records with activity from more than 500,000 GitHub developers. Across successive tool generations, the paper’s headline comparison estimates a cumulative increase of about 240% for commits, but 30% for releases.
A commit records a code change. A release packages a software version for distribution. Neither alone measures usefulness. The paper cannot isolate autonomous agents’ effect on releases, which are recorded for whole repositories.
The researchers match adopters with developers who had similar prior activity, using earlier control periods to reduce contamination from AI use. This is observational research, not a randomized trial. The gap is consistent with constraints elsewhere in production; it does not prove that human code review caused the entire difference. Two authors disclose paid research consulting for Microsoft.
The practical question is therefore where an extra agent-hour goes: into another draft, a checked change, or a product someone can use. That is the distinction behind asking what counts as done and redesigning a workflow rather than just learning a tool.
An organization can make its machines busier without removing the step that limits delivery. The useful gain is not another hour on the agent clock. It is more finished work passing through the whole chain.
