Linking Code and Prose for Coding Agents

Summary & Methodology Analysis

The notlob system addresses the fragmentation between documentation and implementation by co-locating prose and code within a single artifact. It utilizes a name-graph data structure that maps concept names to their corresponding code blocks and prose. This graph allows machine agents to perform navigation and consistency checks, functioning similarly to IDE shortcuts that resolve type declarations or identify function callers. The architecture relies on a Lark grammar to handle the parsing of these mixed-language artifacts. By maintaining this structured mapping, the system aims to provide coding agents with higher quality context than traditional file-based organization allows. The authors implemented language-specific bindings for Haskell, Python, and TypeScript to enable the system to bridge the gap between high-level specifications and executable code. The development of the notlob implementation and example projects was conducted using Claude Code, with the Sonnet and Opus models assisting in the construction process. This approach is intended to maximize the utility of information within the context windows of large language models (a deep learning model that predicts the next token in a sequence to generate text). However, the paper explicitly notes that experimental evaluation of this design is left for future work. A significant limitation identified by the authors involves platform-specific tooling. Development tools integrated into agents, such as grep commands used within Claude Code, frequently employ optimizations that can undermine the file-adjacency benefits of the notlob structure. These tools may restrict output in ways that degrade the effectiveness of the system for machine agents.