amplifying.ai
|
ksl
|
|
A study of 2,430 Claude Code sessions across Sonnet 4.5, Opus 4.5, and Opus 4.6 found the agent consistently builds custom implementations rather than recommending commercial tools – writing JWT auth from scratch instead of suggesting Auth0, rolling config-based feature flags over LaunchDarkly. When it does pick a tool, the preference is sharp: Vercel owns 100% of JS frontend deploys, Zustand beats Redux entirely, and Vitest has effectively replaced Jest. Opus 4.6 skews noticeably toward newer libraries like Drizzle while Sonnet 4.5 stays conventional with Prisma and Redis. The bigger question is what happens to tool discovery when a single agent’s defaults start shaping how thousands of projects get scaffolded.
