Optimizing Sequential Multi-Step Tasks with Parallel LLM Agents