Grounded Skill Synthesis from Code at Scale for Agentic Intelligence

Grounded Skill Synthesis from Code at Scale for Agentic Intelligence. By Yongqi Tong and six others, from Ant International. Published on the fourth of September. Its code repository is public and real: about twenty source files.

Reusable skills give agents procedural knowledge they can carry from one task to the next, so acquiring skills at scale matters if agents are to reach beyond what they have already experienced. The two existing routes each carry a limitation. Synthesizing skills from trajectories requires interacting with specific environments, so the agent must already have been somewhere before it can learn anything. Deriving skills from documents avoids that, but such skills may lack executable evidence and verification. One route needs experience first. The other may not be checkable at all.

Source code offers a complementary path. It requires no prior agent experience, and it still provides executable evidence in which to ground an abstraction. Code2Skill is a fully automated pipeline built on that observation. Rather than watching an agent act, or reading documentation, it reads code and turns it into skills.

The pipeline takes selected code units and transforms them into records anchored in their implementation. The records come in three kinds: atomic operations, composite workflows, and recurring patterns. Each record is then verified twice over, through reconstruction that is blind to the source body, and through comparison that is aware of the source. Applied to nineteen thousand seven hundred and sixty-nine popular, actively maintained GitHub repositories, the pipeline produces CodeSkillBank. It is a grounded bank of one million six thousand eight hundred and twenty-two accepted records, each carrying metadata on workflow, boundary, provenance, and source evidence.

The evaluation spans seventy-two protocol-matched comparisons, covering nine model settings and eight benchmarks. Models augmented with skills retrieved from CodeSkillBank improve by eleven point seven percent on average over matched baselines, and outperform them in fifty-seven of those cases. Under a unified downstream interface, Code2Skill also beats skill banks derived from trajectories on all seven shared benchmarks. The authors read that as showing repository-derived skills can supply useful procedural knowledge before an agent has gathered enough interaction experience of its own.

A further test looks ahead. Skills synthesized from tested code that was itself generated by AI reach a ninety-three point five percent pass rate, against ninety-three percent for skills from human-written code. The authors present this as initial evidence that the pipeline can keep expanding as the volume of AI-generated software grows.

Put plainly, the bet is on where procedural knowledge already lives. It is embedded in repositories, written and exercised long before any agent arrives, and Code2Skill converts it into skills that are grounded, verifiable, and transferable, without first sending the agent out to gather experience.