I traced the agentic calls. Here's where the token consumption comes from

An analysis of an agentic coding assistant's token consumption reveals a significant difference compared to simpler tools. The workflow involves a user prompt, the LLM fetching file contents, generating internal thoughts on UI design (e.g., adding color blue), and using an edit tool to restructure layouts in A.py and B.py. The Harness confirms successful replacements, and the LLM further edits B.py for CSS classes before summarizing design changes like "Modern layout" and "Visual Hierarchy" to the user.

This report uniquely details the step-by-step token consumption of an agentic coding assistant, unlike general discussions of LLM token usage.