Armature benchmarks coding agents’ tool picks
Armature analyzed 16,893 coding-agent sessions across 75 repositories, 1,163 task variations, and Claude Code, Codex, and Cursor to see which third-party services they actually install. Only 42% of agent decisions matched, with programming language and repository context often changing the winner.
The key finding is that agent discoverability is becoming a distribution channel, but there is no universal “best” tool—context, search behavior, and documentation shape every choice.
- –The same email task favored Resend in TypeScript, SendGrid in Python, Postmark in Go, and Azure ACS in Java.
- –Codex searched the web in 94% of sessions, Cursor in roughly two-thirds, while Claude Code relied more heavily on prior knowledge.
- –Claude Code built solutions in-house nearly twice as often as the other agents, showing meaningful behavioral differences between coding systems.
- –Mentions did not translate into adoption: PayPal was cited 139 times but never selected, while LangChain was mentioned 194 times and picked only four times.
- –The study is especially useful for tool vendors, though its conclusions should be read alongside Armature’s commercial interest in improving agent discoverability.
DISCOVERED
1h ago
2026-09-04
PUBLISHED
4h ago
2026-09-03
RELEVANCE
AUTHOR
screm