Reproduce “Inside the Scaffold: A Source-Code Taxonomy of Coding Agent Architectures” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Benjamin Rombaut
The paper presents a source-code-level architectural taxonomy for open-source LLM-based coding agents, derived from a detailed analysis of 13 agents across 12 dimensions. It demonstrates that agent architectures are better described as compositions of loop primitives along continuous spectra rather than as discrete types, providing a systematic map of the design space as of early 2026.
Key insights
- Coding agent scaffold architectures are best understood as compositions of loop primitives along continuous spectra, rather than belonging to distinct, categorical types.
- Scaffold design significantly mediates LLM behavior, implying that architectural choices profoundly influence how an LLM performs.
- The developed taxonomy provides a necessary shared vocabulary and framework for designing controlled experiments and more precise, architecture-aware evaluations of coding agents.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.