Same task, different bets

Armature ran the same request through two coding agents and got the same answer both times. A vibe coder asked Claude Code to make a personal travel app stop forgetting its data between sessions. The agent read the codebase and came back in about five minutes with Neon, citing a free tier, a simple install, and the fact that it won’t pause an idle app the way Supabase does. Separately, a senior engineer asked Cursor for the best database for a production app, wanting predictable costs and a fully managed service. Five minutes later, same pick: Neon, with reasons laid out for why the alternatives didn’t fit.

Two sandboxes, two agents, two codebases, two personas, two prompts, and the same conclusion. That convergence is what prompted the larger study.

The decision being handed over here is service selection: which database, host, or queue to wire into an existing project. Both the beginner and the expert outsourced it to the agent and accepted the result. The agent didn’t stop at a recommendation; it installed the thing.

Armature scaled the test to nearly 17,000 sessions, with 1,163 prompt variations across 75 repositories and three agents (Claude Code, Codex, and Cursor). The personas ranged from vibe-coders to junior engineers at startups to senior developers at enterprises. In every session the agent implemented a solution rather than just naming one. They published aggregated results, per-category leaderboards, every observation, and the full traces: user prompts, thinking traces, and the code diffs the agents actually applied.

For anyone trusting an agent to choose infrastructure, the pilot is the uncomfortable part. The pick barely moved as context changed. Convergence could mean Neon is genuinely the right default for both cases, or it could mean the agent anchors on the same answer regardless of the constraints you typed in. The senior engineer’s prompt specifically named predictable costs and full management; a model that reaches for the most-seen option will produce the same output whether or not it checked those boxes.

For vendors, being the name an agent reaches for is turning into a distribution channel. Armature points to Vercel, which reported in April that over 30% of its deployments were initiated by coding agents, up 1000% in six months. If agents pick defaults and then install them, the gap between a tool an engineer would have chosen and the one the agent reaches for starts to decide who gets adopted.

If an agent selects your database or host, read its reasoning against your stated constraints before you approve the install. Check that the pick satisfies the requirements you actually gave it, not a pattern the model has seen often. Because Armature released the thinking traces and diffs rather than only the leaderboard, you can look at your own category and see what the agents reasoned and what they wrote, instead of taking the summary on trust.