Found first: a primary source the press has not covered yet.
A paper posted to arXiv on September 2 introduces DisCo, a method that automatically extracts operational knowledge from GitHub repositories and packages it as reusable skills. Applied to a GPT-5.5 agent with model weights unchanged, the resulting skill-equipped agent scores 134.3% higher on MLE-bench than the same agent running without skills, according to the paper.
What the source says
Jianlyu Chen and ten co-authors (institutional affiliations are not listed in the arXiv submission) ran DisCo across 1,000 widely-used ML repositories to produce the AREX-Skill Library: 5,000+ verified skills organized across 20 capability areas and 178 capability families. Skills come in two modes: task-agnostic, which captures broadly applicable patterns from popular repos, and task-oriented, which targets skills at a specific downstream task. With the GPT-5.5 backbone, research harness, and execution budget held fixed, the skill-equipped agent scores 134.3% higher on MLE-bench, 34.4% higher on PaperBench, 9.2% higher on FrontierCS, and 14.0% higher on PassNet, all relative to the same agent without skills.
Why it matters
The gains arrive without any change to model weights, which means operational knowledge distilled from real codebases functions as a separable, additive capability layer on top of the base model. A 134.3% lift on MLE-bench from an external library suggests that frontier models are far from capturing what domain-specific procedural knowledge can add. This repositions the primary engineering lever: instead of retraining or fine-tuning the model, practitioners can invest in curating and updating an external skill library. The approach also implies that the library itself becomes a competitive asset, distinct from the model it augments.