Skeleton Coding: Program Structure as a Specification for Agents
Source
arxiv.org
Author
Aditya Golatkar, Alessandro Achille, Renos Zabounidis, Vlad Sobal, Yangruibo Ding, Baishakhi Ray, Jonathan Buck, Guy Van den Broeck, Maryam Fazel, Pietro Perona, Wei Xia, Stefano Soatto
Date
Why it matters
Tests alone can underspecify a program because the agent picks the structure that determines which tests suffice. Specifying structure up front gives engineers a way to constrain coding agents beyond adding more tests.
Key takeaways · AI-distilled
Skeleton Coding has developers specify the structure, data types, modules and functions in a high-level code-like form, alongside behavioral specs. Modeling AI-assisted coding as a control problem, the authors show this can cut required tests exponentially.
Across four models and three benchmarks, the best skeleton specifications reduced mean test error by 48.4% relative to the best end-to-end behavioral specifications, and by up to 27.0% even when the AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → could query a complete reference implementation.
Skeletons also narrowed the average gap between small and large models by about 52.4%, lowered implementation cost by 15 to 26 times, and reduced code erosion by 24.7% on average, according to the abstract.
Terms in this piece · Glossary
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
articleCode Understanding is a Bottleneck for Coding AgentsNishant Balepur, Kiran Tomlinson, Tobias Schnabel
articleSpecification-first convergence with an AI coding agent: a case study of dismantling a core architectural invariant across 189 files in a 717k-line codebase with no test oracle and no human code reviewJoel Abenhaim
articleHumanize: Judgement Engineering for Agentic CodingSihao Liu, Ligeng Zhu, Zijian Zhang, Dongyun Zou, Zhengyang Zhang, Changye Li, Song Bian, Song Han, Tony Nowatzki