Give agents a sense of elapsed time and budget — models derive surprising hacks

realsohamparekh · x · 2026-11-01

Physera AI shares practical fixes for agents' lack of time sense: keep models informed of elapsed time and add a budget, and they derive surprising on-the-fly optimization hacks (works with token budgets too). The real fix, they argue, is ingraining time efficiency into the RL reward itself — current frontier RL rewards token efficiency, future RL will reward time efficiency.

Original post →

Give agents a sense of elapsed time and budget — models derive surprising hacks

More from coding & agent