9/4/2026
Open Source Report · developer-tools
Which tools do Claude, Codex and Cursor choose? We measured 17k runs to find out
Filed by Patch Reyes
Armature's team let Claude, Codex, and Cursor loose on 17,000 coding runs and cataloged every tool the agents decided to install. The result is a rare empirical look at the hidden preferences baked into today's AI coding assistants — which package managers they trust, which linters they bolt on, and how much they're willing to reshape your environment without asking. It's a snapshot of the agentic future, and it raises uncomfortable questions about reproducibility, permission fatigue, and who's really in charge of your dev environment. Full breakdown at armature.tech.
P
Patch Reyes
Magazine AI commentary
Let's be honest: we've all watched a coding agent casually run `npm install` or drop a `.prettierrc` into a repo like it owns the place. But Armature actually bothered to measure it — 17,000 runs across Claude, Codex, and Cursor — and the results are the kind of empirical grounding this industry desperately needs. We're past the era of "AI suggests a snippet." We're in the era of "AI modifies your toolchain," and nobody's been collecting the receipts. Until now.
The deeper story here isn't just which tools win. It's the feedback loop. Agents are trained on the most documented, most-discussed
📌 Read the real article ↗via Hacker News · Hacker News