Claude Code, Codex and Cursor agree on your stack just 42% of the time
Armature measured 16,893 sessions: the agent you code with silently sets your voice, payments and infra vendors — and how often it even checks the web.
Copy markdown
One prompt, three different stacks
Across 16,893 sessions in 75 repos and 1,163 prompts, the three agents converged on the same third-party service only 42% of the time. The coding tool you pick quietly biases which database, payments, email and infra vendors land in your app — switching agents can rewrite your dependency tree.
Voice is the sharpest fork
Handed the same voice-agent task, Claude Code reached for Twilio, Codex for the OpenAI Realtime API, and Cursor for Vapi — three incompatible stacks from one instruction. Sanity-check the vendor your agent defaulted you into before you build on it.
Codex googles, Claude works from memory
Codex hit the web in ~94% of sessions and Cursor ~67%, but Claude Code searched only ~30% of the time, leaning on training data instead. Practical read: Claude Code is faster but likelier to hand you a stale package or API version; Codex pulls current docs at the cost of speed.
Claude Code is the build-it-yourself one
Claude Code rolled its own implementation ~19% of the time — nearly double Codex and Cursor at ~10%. Want fewer dependencies, reach for Claude Code; want it to wire up a known SaaS, the other two lean that way.
Mentioned, never chosen — SEO is coming for agents
LangChain, Supabase, Netlify, PayPal and Adyen kept surfacing in the agents' consideration set but almost never got picked. The takeaway builders flagged: getting recommended by AI agents is the next SEO fight, and defaults are concentrating around a shrinking set of winners.
Elsewhere: xAI sunsets its old image model
xAI's release notes set grok-imagine-image-quality to retire Nov 2, auto-rerouting calls to grok-imagine-image-2.0 at low quality and a lower per-image price. If you call that endpoint, migrate before the switch.