One underappreciated reason for the agentic gap: Gemini tends to over-explain its reasoning mid-tool-call in a way that breaks structured output expectations. Claude and GPT-4o have both gotten better at treating tool calls as first-class operations. Gemini still feels like it's narrating its way through them rather than just executing.