handwrite your user level AGENTS.md that holds your preferences, and treat your project level ones like a neural net. Give it a size budget and train it on the sessions that actually ran.
Your AGENTS.md is a Neural Net · Kun Chen · 23 August 2026
If you put all these things together:- RLHF = training the model to be likable by humans- RLVR = training the model to be accepted by machines- RLVR is more scalable- "Alignment tax" says "likable by humans" makes the model do worse on verifiable tasks
long-horizon agents assume requirements all exist upfront. They are fundamentally against human in the loop. And they don't have true judgment for what humans like. That—is why I don't like them.