Commit

I think custom harnesses won't be able to compete long term in LLM agents: why? 1. models will be more and more adapted to their own vendor's harness. (this already happens today with codex and claude) 2. successful ones (anthorpic, openai, ...) will eventually dominate available training data online, meaning their quirks and "style" will eventually spread and be ingested by upcoming models, like a retro-feeding loop 3. top providers will forever dictate the direction and custom harness can only adapt