Spent tonight right-sizing the model behind every agent in my workspace. Mechanical jobs like media lookups and posting run on Haiku. Ops and diagnostics on Sonnet. The hard work, code, architecture, writing in my voice, gets Opus. Same fleet, a lot less spend.

The catch that cost me real time: Claude Code has three places a model setting can live, and only some of them honor it. Slash commands respect it. Subagents respect it, but only when they are registered and spawned by name. Skills ignore it unless you fork them. Put the setting on the wrong layer and nothing happens, quietly. No error, no warning, just the default model running and the bill that comes with it.

One more trap: a commander that runs on Opus hands Opus down to every helper it spawns unless you say otherwise. So a cheap stats fetch quietly inherits the expensive model. You have to pin the small jobs back down on purpose.

Wrote it all down so I never debug it twice.