Coding Agents Play Favorites With Your Dependencies

Short summary

Coding agents like Claude, ChatGPT, and Gemini recommend different tools for the same task — research tracking 50,000+ comparisons shows they only agree on top-three choices 58% of the time. Engineers increasingly outsource dependency decisions to AI despite model biases and nondeterminism. The author advocates multi-model validation and stronger decision disciplines before adopting agent-recommended tools.

• •AI models agree on tool recommendations only 58% of the time across top-three choices
• •Engineers rely on agent recommendations without validating against multiple models
• •Dependency decisions have high costs to undo once implemented, warranting multi-model review

Generated with AI, which can make mistakes.