I switched my main model and started cutting my agent setup
Swapping the model on my main agent was the cheap part. I moved it to Opus 5.5 as a test, and four days later I wrote this:
opus 5.5 we keep now, test is good
No benchmark behind that. I used it for four days and it worked.
A test I could undo
My main agent in Claude Code plans, hands jobs to helper agents and checks what comes back. It was running on Fable. On September 22 I switched it to Opus 5.5 on high, on my Mac and on my VPS, and called it temporary. The coding and frontend helpers went to Opus 5.5 on medium, with high when the job is more complex. The Sonnet and Haiku helpers stayed as they were.
It was easy to reverse, which is why I was happy to try it. I judge it by the code coming out cleaner. I like how it’s working now.
High stays on the orchestrator
The main agent orchestrates, so I want it thinking. I did ask it to tell me if that’s a bad idea and to suggest running orchestration on medium instead. It can check how I actually use it through my mentor skill, which keeps most of that information.
The setup had grown past the model
The model was never the heavy part. The pile around it was.
On August 26 I sent scouts out to research how to make skills load on demand instead of always, and how to keep CLAUDE.md and the other files clean. I wanted to stay away from the usual LLM way of doing it and give the agent only the tools it needs. Then I looked at what I had:
39 skills maybe stil la lot
That day some skills went on demand only, like the video frame ones. The brand and marketing writing skills got switched off, and I kept just the naming one.
On September 18 the mentor skill went through how I code with agents and showed where context was bloating and where things got more complicated than they needed to be. I doubted my CLAUDE.md did better than ponytail, a plugin that pushes the agent toward the smallest change that works. So I wanted ponytail across the whole setup, and every .md file made clean and concise, keeping everything but saying it shorter. Then:
claude.md 9kb is a lot, what is there? can make it more concise?
The same day I dropped a media generation tool, because I use my media canvas app now. Ponytail stayed. The September 26 review of my Claude, Codex and Cursor setups went after the same things: cleaner code, better decisions, less context bloat and fewer tokens. It flagged the CLAUDE.md for ponytail compliance and the marketing skill list for trimming. The hyperframes skill stays.
Cursor needs the harness most
I also use Cursor with Grok 4.6. It’s fast, but it rushes decisions, so it needs ponytail and the rest of the harness around it more than any other model I use. I’m keeping Grok in Cursor.
Did ponytail make a difference? The mentor skill can compare before and after. My own read is that the code comes out cleaner.
I haven’t gone through the skill list properly yet. I want to keep the few I reuse most, like impeccable and ponytail, and I’m not sure which others I open often.