Rigor Compresses: Why AI Agents Need Graphs, Not More Context

Short summary

AI agents struggle with global code questions not from context limitations but from guessing at facts they can't verify. Rather than bigger windows, provide verified facts extracted as graphs: dependency graphs, control flow, reachability analysis. Using tree-sitter for quick parsing and LSP for verification, you hand agents verified signal per token, replacing confident guesses about impact and safety.

• •Agents excel locally but guess at global facts like dependencies, impact, and ordering
• •The solution is verified facts formatted as graphs, not larger context windows
• •Use tree-sitter and LSP to produce cheap, verified facts that replace confident guessing

Generated with AI, which can make mistakes.