Architecture Descriptors Cut AI Coding Agent Navigation by 33-44%: 7,012 Claude Code Sessions Analyzed

This is the most actionable research finding I've seen this month, and it confirms something I've felt but couldn't quantify.

Paper arXiv:2604.13108 studied 7,012 Claude Code sessions and found that structured architecture documents, ones that declare module boundaries, symbol signatures, and data flows, reduce agent navigation steps by 33-44%. The statistical evidence is unusually strong: Wilcoxon p=0.009, Cohen's d=0.92, with a 52% reduction in agent behavioral variance.

The proposed format is called intent.lisp: an S-expression syntax where projects declare their architecture for agent consumption. Automatically generated descriptors achieved 100% accuracy compared to 80% when agents navigated blind.

This challenges a popular assumption. Many builders assume coding agents should figure out codebases on their own. Just point Claude Code at the repo and let it explore. The data says that's leaving 33-44% of your agent's efficiency on the table. When I look at my own CLAUDE.md files and architecture docs, this tracks. The sessions where I've pre-loaded context about module boundaries and file locations are noticeably faster. I just didn't have numbers for it until now.

The practical takeaway is immediate. If you're using Claude Code, Codex, or any coding agent daily: write a structured architecture document that declares your module boundaries, key symbols, and data flow patterns. Not a README. Not inline comments. A dedicated machine-readable document that tells the agent where things are and how they connect. The paper suggests the format matters less than the presence. Even a well-structured markdown file with clear headings for each module will help.

52% reduction in behavioral variance is the number that excites me most. It means your agent's performance becomes more predictable, not just faster. Less "sometimes it finds the right file in 2 steps, sometimes it takes 15." More consistent results, every session.