What's the biggest token waste in your AI coding workflow?

I've been benchmarking AI coding agents for the past few months and one finding keeps coming up: agents spend ~80% of their context budget on orientation, reading files, grepping, exploring - before they write a single line of code.

On a FastAPI codebase (~800 files), Claude Code averaged 23 tool calls per task just to figure out what's relevant. That's 40K+ tokens burned before any actual work happens.

I've been building a solution to this and we're launching it on Product Hunt tomorrow. But before that, I'm curious:

What's the biggest source of token waste in your workflow? Is it the exploration phase, context window exhaustion mid-task, session restarts, or something else entirely?