Armature.tech ran 1,163 prompt variations across 75 repositories through three coding agents, Claude Code, Codex and Cursor, and watched what each one actually built rather than what it recommended. The company describes the method on its blog: nearly 17,000 sessions across personas ranging from "vibe-coders" to "junior engineers in startups" to "senior developers at enterprises," with agents "actually implementing the solutions instead of just recommending one."

The headline finding is agreement, or the lack of it. All three agents choose the same tool in only 42% of the cells the study measured. In the voice agents category, Claude Code reaches for Twilio, a phone and messaging API. Codex picks the OpenAI Realtime API instead. Cursor goes with Vapi, a separate voice AI platform. Same prompt category, three different vendors built into the code.

The agents also diverge on habits, not just picks. Codex uses web search in 94% of sessions, and in the page's own account, nine times out of ten it narrows that search with operators like `site:`, aiming at a specific trusted domain rather than searching broadly. Claude Code, meanwhile, builds functionality in-house nearly twice as often as the other two, at 19% of sessions against 10% for Codex and Cursor.

That gap sits awkwardly next to how these tools are usually sold: as interchangeable front ends where the model is the choice that matters and the agent wrapped around it is a preference of taste. The measured behavior says otherwise. An agent that defaults to building in-house produces a different dependency list, and a different maintenance burden, than one that defaults to pulling in a third-party API.

For a builder picking a stack, that turns "which agent do I like" into a question with product consequences. A team using Claude Code for a voice feature ends up wired to Twilio's pricing and API surface; the same prompt through Cursor ends up wired to Vapi's instead. Neither choice is visible until the code is already written, unless someone runs the comparison first.

The study does not say why the agents diverge this sharply, whether it comes from training data, system prompts, or something else built into each product. It also does not say whether the tool preferences hold steady as each vendor updates its agent, or whether a Claude Code release next month reaches for a different voice API entirely. That is the open question for anyone building a stack on the assumption that today's default is next quarter's default too.