Claude doesn't choose tools—it hallucinates JSON instead. When you send a message to Anthropic's API with a list of available tools, the model doesn't "think" about which one to pick in any human sense. Instead, it runs a structured, multi-stage pipeline that converts natural language intent into rigid, machine-parseable tool invocations—and the entire process happens in under 200 milliseconds.
Read the full article: https://hubaiasia.com/how-claude-decides-what-tool-to-call/
Top comments (0)