What it costs to re-send an agent's context every loop

Tokens consumed on a single heading edit — before and after we stopped repeating the context each turn

Running code costs something — and you can measure it. On one real task, more than half the tokens were spent re-sending context the agent already had. The numbers live on a public issue, where anyone can re-run them: AI Context #3582288.

Tokens consumed on one heading-edit task, by approach
Approach	Tokens
Re-send full context every loop	101,000
Send only what changed	48,000