Armature announced that it has investigated the process by which coding agents select which services to implement for specific needs.
NewsClaude CodeCodexCursorArmature
Large-scale survey results released on development tools selected by coding agents such as Claude and Cursor
This article is a translation. Read the Japanese original
In this experiment, three coding agents (Claude Code, Codex, and Cursor) were used. Using these, nearly 17,000 sessions were conducted under various personas, prompts, and codebase conditions.
The study utilized 75 repositories and 1,163 patterns of prompt variables. In the experiment, the agents were not merely asked to recommend tools, but were tasked with actually implementing the solutions.
As a result, detailed data has been compiled, including leaderboards by category, the agents' thought processes, and the actual code diffs applied.