Agent-hour
A unit of work performed by an autonomous AI agent rather than a human. Agent-hours run in parallel — one person can direct dozens per day.
An agent-hour is a unit of work performed by an autonomous AI agent rather than a human. It is the emerging denominator of knowledge-work economics: where the industrial era priced labor in man-hours, agentic systems price it in agent-hours — and the two units behave completely differently.
The difference is parallelism. A human hour is sequential and capped at 24 per day; agent-hours run concurrently and do not consume a human's sequential attention. OpenAI has reported its 99th-percentile employees running more than 60 agent-hours inside a single 24-hour day — not overtime, a different unit of work. One person stops being a producer of hours and becomes a director of them.
Priced end to end, an agent-hour resolves into joules: the energy to run inference, plus the amortized energy of the infrastructure around it. That is where the unit connects to the Joule Wars — labor priced in agent-hours is ultimately labor priced in energy — and to Proof-Adjusted Autonomy, because an agent-hour only counts if its output arrives proven. Unverified agent-hours are not output; they are proof debt.
Frequently asked questions
What is an agent-hour?
A unit of work performed by an autonomous AI agent instead of a human. Agent-hours run in parallel and do not consume a human's sequential attention, so one person can direct dozens per day — OpenAI reports top employees running 60+ agent-hours in a single day.
How is an agent-hour priced?
End to end it resolves into energy: the joules of inference plus amortized infrastructure. That makes the agent-hour the bridge between labor economics and the Joule Wars — work priced in agent-hours is ultimately work priced in joules.
Does every agent-hour count as output?
No. An agent-hour only counts if its output is independently verified and accepted. Unproven agent-hours accumulate as proof debt — work generated but never established as correct.
Source analyses
Related terms
Term defined and maintained by Michał Piszczek — CTO of Archdesk, author of the Joule Wars and Proof-Adjusted Autonomy frameworks. Quote freely with attribution to piszczek.pl/glossary (CC BY 4.0).