Tag
A study finds that 58% of tokens consumed in failed AI agent runs occur after the model had a clear signal to stop, highlighting a missed opportunity for mid-run cost savings.