Fixed overhead first
System prompt plus tool definitions are paid on every request before anything useful is loaded. On a heavily-tooled agent that can be several percent of the window, permanently.
Seeing it as a share of the window rather than as an absolute number is what makes people trim it.
Why the safety margin is not optional
Token counts for models without a public browser tokenizer are estimates, and a request that exceeds the window fails outright rather than degrading gracefully.
Fifteen percent of headroom costs nothing you will notice and prevents the class of failure that only appears under production inputs.