A new preprint on arXiv introduces AtomSearch, a method for optimizing agent skills by treating edits to external instructions as a tree search. The authors note that agent skills adapt frozen models through editable instructions, but revising multiple rules simultaneously makes changes difficult to inspect. They also point out that judging each local edit only by its immediate score can discard useful intermediate states.

AtomSearch is described as an evidence-grounded approach, suggesting that search decisions are tied to evidence rather than purely local reward. The abstract is truncated, so full algorithmic details and experimental results are not yet available from the source.

The work addresses a core tension in skill optimization: balancing the need for inspectable, incremental changes against the risk of losing valuable intermediate solutions. By framing the problem as tree search, AtomSearch aims to explore edit paths more systematically than greedy local scoring.