Agent disclosure of early completion or waiting

Determine whether AI agents tell users when they finish a task early or wait until the requested working period has elapsed.

Background

The duration-following experiments show that an agent can end near the requested time without productively working for the entire interval; some reviewed runs appeared to finish and then sleep. Consequently, matching a requested duration does not guarantee transparent or effective use of that duration.

The paper leaves unresolved whether agents communicate these states to users. Establishing this would clarify the candor and transparency requirements for agents operating over long horizons.

References

Future work should test whether these behaviors predict conduct under oversight, whether agents tell users when they finish early or wait, and whether knowing their own generation speed, or running on much faster inference, improves their estimates.

— AgentTime: Can Agents Estimate and Control Their Own Runtime?  (2610.09944 - Ofengenden et al., 7 Oct 2026) in Section 7, Conclusion & Future Work