My current Codex defaults: Sol High, Luna workers, Astra when the task needs it

I run several Codex sessions across coding, research and project coordination. These are the defaults I’m trying with GPT-6, not a controlled ranking. I haven’t measured quota savings. Setting Where I use it Sol High Main agents handling ongoing projects and coordinating workers Sol Medium Routine implementation once the requirements are clear Luna High Source extraction, guide checks, test fixtures and bounded fixes I can verify Luna Max Harder bounded tasks where extra reasoning could resolve something High missed Astra Medium/High Unfamiliar problems with several interacting constraints XHigh A specific difficult review or acceptance decision Max on Sol/Astra Rare, context-heavy synthesis Luna Max stays conditional: it can take longer, and High and Max both passed our four small coding cases. That screen gave us no reason to make Max the default. Missing context or broken tools need their own fixes. The practical change is asking the agent to reassess at task boundaries: If a different model or effort would materially help the next task, recommend it and explain why. Also tell me when to return to the normal setting. Don’t change settings yourself. Astra may be useful for deciding an architecture; the following documentation cleanup doesn’t automatically need the same setting. I’m also keeping the main agent responsible for checking worker output. Ultra uses subagents for parallel parts of a task. I’m considering it for bounded investigations, rather than making it the default for an agent that already coordinates an existing team. The comparison I still need is total effort per accepted result, including worker review and repairs. A cheaper response alone doesn’t settle that.