IME, agents are generally capable of passing the tests autonomously. But not with code you'd actually ever want to use. Both because of the quality, and because it will have awful performance/memory/security issues.