AGENTS.md for Reconstructing a Tiny GenAI Library: Initial Analysis of Trivial PyPI Packages

Keywords: AGENTS.md, Empirical Study, Generative AI, Software Libraries

TL;DR: Reconstructing PyPI trivial libraries using generative AI with AGENTS.md as specification

Abstract: The rise of agentic systems offers a dynamic alternative to traditional third-party library reuse, promising on-the-fly functionality adaptation. Central to this shift is AGENTS.md, which serves as the contextual guide for navigating Generative Artificial Intelligence (GenAI) environments. In this new ideas paper, we explore how libraries could be generated on-demand through deconstructing the specification contexts in AGENTS.md. Our study involves generating 15 different combinations of AGENTS.md to generate 5,760 specifications and 80 GenAI libraries from 40 randomly sampled libraries. Our initial analysis reveals that context is critical: function descriptions and implementations account for the most of generated documentation, while testing and overviews generate the least. With the growth of AI Slop, we empirically show that specifications have potential to guide GenAI for tiny code compared to human-written equivalents. We finally present a research agenda for efficient, on-demand GenAI library technologies.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public.

Paper Type: Short papers (i.e., vision, new ideas, and position papers). 2–4 pages

Reroute: false

Submission Number: 57