Where do coding agent "skills" come from? Two routes have dominated: let agents accumulate skills through real interactions, or extract procedural knowledge from documentation. The first is expensive — every skill costs environment interaction; the second lacks executable evidence and is hard to verify. A paper published on arXiv on Sep 4 (arXiv:2609.05571) from Ant International proposes a third route: synthesize skills directly from source code. As of the Sep 21 Hugging Face Daily Papers board it ranks #1 of the day with 86 upvotes, submitted by the first author.
Method: source-body-blind regeneration
The Code2Skill pipeline runs in seven steps: it first scores parsed source units — functions, methods, command entry points, file-level components — for reusable procedural evidence; then extracts typed skill records in three classes: atomic operations, composite workflows, and recurring patterns. The core validation step is "source-body-blind regeneration": a model receives only the skill description and interface — not the source body, repository name, or file path — and must regenerate the code from that alone. A separate equivalence judgment compares source and regenerated code, and an adjudication stage separates "the skill is unsupported" from "the regeneration failed", so good skills are not discarded by accident. Only records passing everything enter the bank, tagged with workflow, boundary, provenance, and source-evidence metadata.
Numbers: a million records, 72 matched evaluations
The reported scale is substantial: 19,769 popular, actively maintained GitHub repositories yield 1,006,822 accepted records. With retrieved skills attached, models improve by 11.7% on average across 72 protocol-matched evaluations spanning nine model settings and eight benchmarks, winning 57 of them; against trajectory-derived skill banks, code-derived skills win on all seven shared benchmarks — meaning repository-mined skills help before an agent has accumulated interaction experience. One more striking pair: skills synthesized from AI-generated code pass at 93.50%, versus 93.00% for human-written code, nearly identical. As AI-generated code keeps growing, this pipeline can scale with it rather than run out of material.
Caveats: the judge is not a proof, and the code is not open yet
The GitHub README is blunt: the judge "is an LLM consistency check, not a proof of program equivalence." The million-record "accepted" standard is still model-decided, not mathematically verified. The repo reality: it is a pre-release with only 3 commits and no license attached; the README states plainly that it "is not yet a public open-source release." What is actually open is the companion dataset, DeveloperSkills-Code2Skill on Hugging Face — the first author shared the link in the paper's comment section. Anyone wanting to reproduce the full pipeline will have to wait.
Why it matters
This work reframes skill acquisition from accumulated experience to asset mining: the code already sitting on GitHub is the largest existing carrier of human procedural knowledge, and there is now a batch extraction path that is verifiable — albeit verified by an LLM. For Ant International, it is another agentic-infrastructure bet after the Ling series of agent models. Where the LLM-judge reliability ceiling sits, though, will only be answered when third parties reproduce it inside real agent workflows — the same open question hanging over every "synthetic data + model self-verification" pipeline.
References: arXiv:2609.05571 (arxiv.org/abs/2609.05571); GitHub ant-intl/Code2Skill; dataset huggingface.co/datasets/ant-intl/DeveloperSkills-Code2Skill