Preprint decomposes why agent skill libraries help and hurt, on office automation benchmarks

Sources: arXiv 2607.22520
Summary: A preprint titled The Regression Tax decomposes why skill and instruction libraries help and hurt LLM agents. The comparison corpus is nearly 6,000 runs with and without skills, spanning two office automation benchmarks and three model harness stacks, not coding-agent work. It reports that the best performing skills outperform others primarily by regressing less rather than by gaining more, and names three mechanisms it calls description osmosis, grounding displacement, and verification displacement. This is a single preprint, not independently reproduced, and the run count and failure taxonomy are the authors' own.
Why it matters: Skill libraries are usually assumed to be free to add, and a named set of regression mechanisms gives teams something specific to test for, though the measured domain is office automation and transfer to coding agents is not established here.