A new research method scores each step of an AI agent's instructions to find out which ones are actually doing any work.
Researchers built a framework called SkillShapley that uses Shapley values, a game-theory tool for splitting credit among contributors, to measure how much each step in an "agent skill" (the written instructions that let an AI agent complete long tasks like coding or document processing) contributes to whether the task succeeds. The team found that naive scoring does not work well here for two reasons: benchmark scores tend to jump in sharp cliffs rather than move smoothly, and most steps interact additively rather than compounding each other's effects. So SkillShapley runs in two phases, first hunting for the step combinations that actually produce useful signal, then sampling new combinations designed to reuse that evidence instead of testing every possible subset from scratch. Tested on skills from SkillsBench, a benchmark suite for these agent instructions, the method reliably flagged which steps were pulling their weight and which were not.
Most agent skills today are still written by hand or pulled from an agent's own execution logs, with nobody checking whether individual instructions actually help. A tool that can cheaply flag useless or harmful steps means skill authors could prune bloated instructions based on evidence instead of guesswork, which matters more as products increasingly ship with these instruction sets baked in.
It is still a research paper tested on one benchmark, not a tool you can install - but it is a rare bit of accountability in a field that mostly ships agent instructions on faith.