One evolution recipe across eight languages: the harness playbook transfers, the ecosystem machinery doesn't
arXiv 2608.10178·low signal
Researchers applied a fixed self-evolution recipe across eight languages and three models to see what evolved harnesses actually encode. The evolved harnesses converge on a shared abstract playbook, but each instantiates language-specific machinery, and while the universal core transfers cleanly the ecosystem-specific components resist generalization. For anyone maintaining agent harnesses across a polyglot codebase, this predicts which parts of a working setup are worth porting and which need to be re-derived per language.