Claude Code, Codex and Cursor chose the same tool in only 42% of the grouped comparisons in Armature’s September 3 study. Repository context and each agent’s search habits changed which services it installed.
Armature ran 16,893 sessions across 75 synthetic repositories in ten languages. Its first published set contained 5,292 sessions on 51 codebases across 18 sectors. Gemini 3.7 Flash played a simulated user and, in a separate instance, judged whether sessions were valid and which tools were chosen.
Codex searched the web in 94% of sessions, compared with about 30% for Claude Code and two-thirds for Cursor. When Claude Code did search, it visited three times as many pages as Codex. The same email-integration request produced four different winning providers across repositories in four programming languages.
Being mentioned was not the same as being selected: PayPal appeared in 139 payment-provider conversations without winning any of them. These are results from Armature’s controlled tasks, not measurements of real customers’ purchases. Armature sells growth services to developer-tool companies and says the study is part of its work on influencing agent choices.