tl;dr: Armature ran 17k sessions (5.3k valid) across Claude Code, Codex, and Cursor to see which third-party tools coding agents pick when implementing features. The agents rely on different sources (Codex heavily web-searches, Claude Code uses priors and builds in-house ~2x more) and agree on the same tool only 42% of the time, with repo language heavily influencing choices—e.g., Resend wins for TypeScript email, Sendgrid for Python. Some categories are lopsided (Stripe wins 90%, Neon 66%, S3 45%), while heavily-mentioned brands like PayPal, LangChain, and Supabase rarely get picked.