This work investigates whether collaborative LLM agents can automate reinforcement-learning post-training for zero-shot text-to-speech. AgenticTTS-Forge organizes human guidance and agent execution around a shared workspace, recovers and improves an underspecified training recipe, and reveals failure modes involving hidden data leakage, proxy inflation, and non-composable policies. The results show that reliable measurement contracts, rather than reasoning capability alone, are central to safe and effective autonomous research workflows.