I've regained my zen by switching back to @LucenaCoder with GLM 5.2 It just works. You don't get piles of words for the sake of words, it delivers for absurdly cheap on that model specifically for some reason, and it's right on the first try rather than defaulting to lazy brittle choices. Also, I realized how much I was fighting audio transcription before (I tried the new grok 4.7/cursor setup) and it's night and day difference. This thing doesn't miss a word, and tells the LLM it's voice transcribed so it doesnt' get myopic and start focusing on mistranscribed things as new directions. Like my calm level returned after switching, it was massively noticeable. I'd rather pay a la carte for GLM tokens via OpenRouter, it's cheaper because it's done right the first time. What's interesting is how much differently Gemini Flash 2.8 behaves there. I'll use it, but controlled only for design or copy tasks otherwise it just endlessly wants to "stay quiet and work" and surfaces later with a huge number of tokens used relative to what might have been used by GLM 5.2. Highly recommended if you're tired of screaming at the mainstream harnesses wanting to sit on a fence and pour piles of word slop at you.

There will always be conflict between those making harnesses and those making models. The labs want the models to be more intelligent and more proactive and burn tokens on the way generating revenue. The harnesses want the models to respond to architecture and not brute force through everything while blowing tokens. Literally every new model release gets a bit more aggressive in how far it will go, which changes how a harness must respond to that model if it cares to keep it contained and not a money pit. The RL training leans on "add" and "solve more" so harnesses are left to fight this behavior as it impacts their interaction quality. It's weird to be involved in. This weird "what IS best?" time. This is where MCP really does win. You can let anyone use whatever model they want for the MCP call, but build your own intelligence behind the MCP that doesn't depend on flavor of the month. You may even pin an older model with RL training that does the job perfectly. Specialists, not generalists. They do one thing well, and that's the entire MCP's job. The user's model then can't really negatively impact it. MCPs should deliver outcomes, not tools.