Why did my agent call the wrong tool with invented arguments?

The model matched your request to whichever tool’s description sounded closest, then, having decided a tool was needed, filled in an argument it did not actually have with a plausible-looking guess instead of stopping to ask. Both failures trace to the same cause: tool calling gives the model no way to say it is unsure, only a schema to fill in, so a low-confidence match and a high-confidence one produce an identical-looking call. Vague or overlapping tool descriptions make the wrong-tool half worse, since the model is choosing on the text of the description, not the code behind it. Missing or ambiguous context earlier in the conversation makes the invented-argument half worse, since the model has a slot to fill and nothing real to put in it.

The fix lives in the tool definitions more than in the model: names and descriptions specific enough that no two tools plausibly match the same request, and required arguments the model can only get from context that is actually present.