Microsoft proposes ParetoSaddler, optimizing agent harnesses for quality and cost

Microsoft proposes ParetoSaddler, a method for optimizing agent harnesses on quality and cost.

Elvis Saravia, summarizing the Microsoft paper, says the method applies the cost preference only at the end of the search rather than at every step, because applying it earlier discards harnesses that later lead to better results.

Saravia reports that on GAIA2, quality rises from 56.22% to 62.89% while cost per task falls from $0.78 to $0.45. On ClawMark, he reports quality rises from 70.86% to 81.35% and cost falls from $1.17 to $1.01.