Found first: a primary source the press has not covered yet.
A new study tests the assumption that skills help AI agents primarily by supplying missing knowledge, and finds that assumption accounts for a small fraction of observed gains. Zhiyuan Jiang and colleagues report in "Demystifying Agent Skills: Why They Work-Until They Don't" that, across 8,135 controlled trials, skills work mainly by anchoring agents into stable procedural sequences. Explicit knowledge injection accounts for 4.5% of successful skill use.
What the source says
The team ran controlled experiments across multiple benchmarks, agent frameworks, and LLMs. They open-coded 240 sampled trajectories, producing 238 valid labels, and analyzed the results to identify when and why skills succeed or fail. Procedural anchoring, in which skills constrain and stabilize action sequences, accounts for 65.7% of successful skill use. Skills outperformed Workflow Memory by 6.06 points in matched comparisons. Retrieval precision drops from 29.6% at 5-item pools to 3.3% at 100-item pools, while downstream task success remains relatively stable.
Why it matters
Skill-library systems are commonly designed on the premise that agents fail because they lack specific knowledge, and that skills fix this by supplying it. This study finds that premise explains 4.5% of the benefit. The dominant mechanism is structural: skills reduce execution variance by constraining what an agent does next, rather than by transferring missing information. The retrieval results carry a practical warning for teams building larger skill pools. Precision collapses from roughly 30% to 3% as the retrieval pool grows from 5 to 100 items, while aggregate task metrics stay relatively flat, meaning retrieval failures can accumulate invisibly as libraries scale.