A thread here on HN last week asked how you know if agents will choose your tool. I dug into the practitioner answers, cross-referenced with Anthropic's engineering data and two recent papers. Turns out it's a deeper problem and this write up goes in real depth.