Rethinking the Value of Agent-Generated Tests for LLM-Based Software Engineering Agents