Large Language Models (LLMs) have become pivotal in advancing the frontiers of versatile agents, yet their effectiveness is still largely reliant on the process of manually crafting prompts, which often remains a complex and resource-intensive challenge. In this study, we present PromptPilot, an innovative framework that addresses the critical challenge of automating the generation of high-quality, task-specific prompts through a novel application of advanced optimization techniques. PromptPilot deploys a fleet of optimizers that independently traverse the vast prompt landscape. Each optimizer refines prompts based on domain-specific feedback and a heuristic evaluation mechanism. Prompt tuning is viewed as a process of state optimization, with transitions between states facilitated by actions that involve sampling and trial-and-error. Consequently, the feedback from error samples facilitates detailed cause analysis and the distillation of experience, leading to deeper domain insights and a better understanding of the task. PromptPilot not only encourages a diverse exploration of the prompt space but also strategically converges on high-quality prompts through a dynamic resampling and branching methodology. Notably, PromptPilot achieves enhanced computational efficiency, with reduced dependency on value function calls. Experimental results across 6 benchmark tasks demonstrate the superiority of PromptPilot over methods like Chain-of-Thought and PromptAgent. By enabling the autonomous generation of precise and optimized prompts, PromptPilot democratizes the utilization of LLMs, paving the way for their broader application across various domains.