We introduce AgentTTS, an LLM-agent framework that efficiently allocates test-time compute across multi-stage tasks, achieving compute-optimal performance through insight-guided adaptive search.