AgentTime: Can Agents Estimate and Control Their Own Runtime?

Michael Ofengenden, Maksym Andriushchenko

arXiv 2610.09944 30 pages Read the paper

The short version

The problem: People who ask an AI agent to work for a set time cannot assume it will respect that request: it may stop early, run too long, or simply wait after finishing. Checking the final answer does not reveal these failures.
The idea: AgentTime tests whether agents can predict how long work will take, estimate afterward how long it took, and work for a requested duration. It checks elapsed time alongside task results and records of the agents’ actions, distinguishing continued work from simply waiting out the clock.
What it got: The strongest agent finished within five percent of the requested time on 63% of runs. Even apparently well-timed runs sometimes included sleeping after the task seemed finished.

Module from this paper

AgentTime: Evaluating Runtime Control

Task scores alone cannot reveal whether an agent can manage its runtime; AgentTime tests duration-following, runtime forecasting, and retrospective time estimation separately.

agentsevaluationtime-awarenessruntime-control