I wouldn't draw such conclusions from one preprint paper. Especially since they measured only success rate, while quite often AGENTS.md exists to improve code quality, which wasn't measured. And even then, the paper concluded that human written AGENTS.md raised success rates.