I turned on auto-approve in Claude Code and broke CI in 30 minutes

Auto-approve looked great until the agent started grading its own homework. Here's what Anthropic's data and real-world drift patterns taught me about designing autonomy.