AIAny
Icon for item

Grounded Skill Synthesis from Code at Scale for Agentic Intelligence

Transforms source code into verifiable, reusable agent skills by extracting atomic operations, workflows, and recurring patterns and validating them via source-body-blind reconstruction. Produces CodeSkillBank (1,006,822 accepted records from 19,769 GitHub repos) and yields ~11.7% average downstream improvement.

Introduction

Most prior work builds procedural skills from agent trajectories or documentation, both of which have limits: trajectories require environment interactions and docs often lack executable evidence. This paper argues that source code is a complementary, high-density reservoir of procedural knowledge that needs no prior agent experience yet provides concrete implementation evidence.

Key Findings
  • Code2Skill pipeline: selects code units (functions, methods, entry points), extracts typed skills (atomic, composite, recurring-pattern), challenges candidate records via source-body-blind reconstruction, and accepts records after source-aware comparison and deduplication. This pipeline emphasizes verifiability over mere summarization.
  • Scale and artifact: applied to 19,769 popular, actively maintained GitHub repositories to produce CodeSkillBank with 1,006,822 accepted skill records, each annotated with workflows, boundaries, provenance, inputs/outputs, invariants, failure modes and source evidence.
  • Empirical impact: across 72 protocol-matched evaluations (nine model settings, eight benchmarks), augmenting models with retrieved CodeSkillBank skills improved average performance by 11.7% over matched baselines and outperformed trajectory-derived skill banks on shared benchmarks. Skills synthesized from tested AI-generated code showed pass rates comparable to human-written code.
Who it's for and tradeoffs

Great fit if you need grounded, implementation-anchored procedural knowledge for coding agents, tool-enabled LLMs, or agent planners and want verifiable evidence tied to real repositories. Look elsewhere if your target behaviors are heavily environment-specific (hardware sensors, closed-system state) or depend on runtime interactions not captured in repository code; the approach abstracts from runtime contexts and may miss deployment-specific nuances.

Where it fits

This work sits between documentation-derived skill summarization and trajectory-derived imitation: it provides transferable procedural building blocks before agents accumulate interaction experience, and it scales with the growth of public (and AI-generated) code.

Methodological notes

The pipeline prioritizes automated verification (blind reconstruction + source-aware comparison) to limit hallucinated or under-specified skills. Evaluation focuses on protocol-matched downstream tasks and cross-benchmark comparisons to measure practical utility rather than only intrinsic metrics.

Information

  • Websitearxiv.org
  • AuthorsYongqi Tong, Pan Wang, Hang Wang, Jianshe Li, Xin Zhang, Jiang-Ming Yang, Wei Wu
  • Published date2026/09/04

More Items

Turns implemented functionality in open-source codebases into executable RL training environments by automatically generating behavioral specifications, execution-grounded tests, and post-rollout filtering; yields 5,545 verifiable coding tasks across 23 languages to train coding agents.

Provides reproducible, multi‑platform environments and execution‑grounded verification for hybrid computer‑use agents that interleave GUI exploration and code/CLI. Key features: a five‑platform harness, RecreationBench (250 tasks), reference‑grounded programmatic and visual assertions for automatic scoring.

Provides a self-evolving ontology layer that enables LLM-based data agents to query and interact with heterogeneous data via an MCP server; it auto-builds and iteratively refines schema, content, and tool layers based on agent interactions.