Context window
The total amount of text a model can hold at once, shared between the conversation, every tool description, and every result returned.
What makes it wrong.
The mechanism, not the warning.
Large result sets are the other half of the problem. A report pulled at daily granularity across a year returns hundreds of rows the model must carry for the rest of the conversation, when the question could have been answered with a weekly aggregate.
In detail.
Tool descriptions are loaded before the first question and stay for the whole session. Results accumulate on top. A window that looked generous at connection can be mostly spent on a catalogue the conversation never uses.
This is why the size of a tool catalogue is a design decision rather than a feature list. Every tool announced is rent paid on every question, including the ones that will never call it.