DEEP_DIVE_STORY.LNK

A real-world test shows multi-model routing slashes AI agent costs, and explicit rules stop agents from quietly deferring work.

In a 2,415-turn log, a model router sent most tasks to cheaper models and cut spend by 94% versus a single frontier model — details and per-model breakdown are in Tyler Folkman’s write-up of his pi router here.
This lines up with reports that many agent steps are deterministic and don’t need an LLM at all; GitHub’s team describes double-digit token reductions by restructuring workflows, covered in Vibe Coding Weekly #32.
One more fix is cultural: coding models often say “we’ll handle this later.” Luna’s analysis ties that to training data, RLHF, and eval blind spots — and recommends explicit “no temporary patches, no deferrals” rules why agents defer.