Study: agent skills stabilize procedures, 65.7% vs 4.5% knowledge

The dominant assumption behind agent skill libraries — that skills work by supplying missing knowledge — does not survive contact with the data. A new arxiv paper from Zhiyuan Jiang and co-authors codes 240 records drawn from 8,135 controlled trials across benchmarks and agent frameworks, retaining 238 valid unique labels, and reports that "procedural_anchor accounts for 65.7% of skill cases, versus 4.5% for explicit knowledge injection."

The mechanism, in the paper's own words: skills work when "noisy trajectories become procedural anchors that stabilize execution."

Against a Workflow Memory baseline, skills gained "+6.06 percentage points, with a 95% bootstrap confidence interval of [+0.76, +11.36]." The lower bound of that interval is close to zero.

The bill comes at scale. As the skill pool expands from 5 to 100 items, the paper reports that "actual-use precision falls from 29.6% to 3.3%" — though downstream success, the authors note, remains stable. The retrieved abstract does not name which agent frameworks were tested or publish a per-benchmark accuracy table.