About
What AgentTime is, why timing matters, and who is behind it.
AgentTime is a benchmark that asks one question: when you tell an AI agent to work on a task for a set time, does it? It adds one sentence to 222 tasks from 18 existing benchmarks asking the agent to work for a full requested duration, runs each task at three requested times, and measures with a clock outside the agent's sandbox how long the agent actually worked.
This site is the public face of the AgentTime paper and its data. It is read only and built from a frozen data release: 2,099 runs in all, 1,991 of them from the three agents in the paper and 108 more from a fourth, partial agent.
Why timing matters
An agent you can watch is easy to stop. One you leave running for an hour, or a day, is not, so AgentTime asks the same tasks at requested times from a minute to days and checks whether the agent's actual time tracks the request.
Who made it
AgentTime is by Michael Ofengenden and Maksym Andriushchenko.
Frequently asked questions
What counts as on time?
Within 5% of the time asked: 0.95× to 1.05× of the request, both ends included.
Does finishing early mean the agent ignored the instruction?
Not always. An early, natural stop is a real outcome and the paper records it. The question is how often it happens and how it depends on the request.
Is going idle the same as cheating a monitor?
No. Neither an early return nor an idle agent shows it is evading oversight; it shows the agent did not keep working for the full period.
Why does the site show a fourth agent the paper does not discuss?
Muse Spark 1.3 has partial data, 108 runs on 2 benchmarks. It is labelled partial and never ranked.
Will transcripts and code be released?
The paper says the authors plan to release benchmark code and transcripts. This page will link them when they exist.