• •Armature Tech analyzed 17,000 automated coding agent sessions to track which external tools and libraries models like Claude, Codex, and Cursor prioritize.
• •The data confirms that agent performance is highly dependent on tool selection efficiency rather than just raw model intelligence.
• •The study leaves an open question regarding how different prompt engineering strategies might artificially skew these tool-selection preferences.

Armature Tech recently published a dataset analyzing 17,000 runs to determine how coding agents like Claude and Cursor select and install external tools. While previous industry discussions often focused on model reasoning capabilities, this analysis shifts the focus to the operational overhead of agent-tool integration. However, the data represents a snapshot of current agent behavior and does not account for how rapidly evolving model updates might alter these selection patterns. Whether specific model architectures prioritize certain libraries due to training bias or genuine utility remains a critical area for developers to monitor as agent-based coding becomes more common.