Coding agents today have a massive spending problem. Every request, whether you’re designing system architecture or writing a single-line docstring, often gets routed to the same expensive frontier model. The result: unnecessary token usage, higher inference costs, and little awareness of task complexity or budget constraints.