Google's Harness² Is This the End of Harness Engineering?

Google's Harness² lets self-improving AI agents rewrite their own harness with zero weights touched. So is this the end of harness engineering? I read the whole paper, appendix included, and the answer is in the part nobody quotes. Google's Harness2, or Harness squared (Google Cloud AI Research, arXiv 2610.14224, open source under Apache 2.0), is recursive agent harnessing: an editor model watches your agent run a task, writes two alternative harnesses, runs both, and keeps only the edits that changed what the agent actually did. No retraining, no answer key. It touches the system prompt, rule