AI agents that rewrite their own harness are overrated, a cranky sysadmin with a shell script still wins
Google's Harness2 pulling 11.5 extra points out of a frozen GLM-5 by letting it rebuild its own scaffolding is not a smarter model, it is a model that finally got a decent desk. Give me one person who actually reads the error log over any recursive agent "learning" its own tooling, it isn't close.