Ant International Releases One Million Skills for Coding Agents

Ant International's research team released Code2Skill, a pipeline that converts code from open GitHub repositories into skills for coding agents, and published the CodeSkillBank dataset with 1,006,822 entries on Hugging Face.

Compiled by Sergey KostenchukPublished 2026-09-22Updated 2026-09-22

2026-09-22 Research Google

🤖 Ant International Releases One Million Skills for Coding Agents

Code2Skill extracts procedural knowledge directly from code in open repositories. The CodeSkillBank dataset is already on Hugging Face: 1,006,822 entries from 19,769 GitHub repositories (500+ stars) — operations, workflows, and patterns verified by code reconstruction.

🌍 Skills from code are available before an agent interacts with a task — unlike Trace2Skill, which requires ready-made successful trajectories. The macro-average across 72 "model × benchmark" pairs increased from 42.90 to 47.90, with improvements on SWE-bench Verified and AIME 2026.

👤 The dataset (619 MB) can be downloaded and used, with cards containing invariants and failure cases serving as a memory base for your own agent. However, the pipeline code is closed (pre-release), and the licenses of the source repositories are not disclosed — verification is required for commercial use.

Source 1: https://ant-international-research.github.io/developer-skill-hubs/ Source 2: https://github.com/ant-intl/Code2Skill

Sources