AUTOACT is an automatic agent learning framework for question-answering tasks that does not rely on large-scale annotated data or synthetic planning trajectories from closed-source models. It enables agents to learn from scratch using a limited set of user-provided data examples and a tool library. The framework starts with a META-AGENT that generates a large database through self-instruct, then synthesizes planning trajectories without external assistance. These trajectories are used to differentiate the META-AGENT into three sub-agents: PLAN-AGENT, TOOL-AGENT, and REFLECT-AGENT, each responsible for task decomposition, tool invocation, and self-reflection, respectively. The division-of-labor strategy allows for efficient and resource-effective learning. Comprehensive experiments with different LLMs show that AUTOACT performs better or on par with various strong baselines. Analysis confirms the effectiveness of the division-of-labor strategy, with AUTOACT's trajectory quality generally outperforming others. AUTOACT achieves self-planning without relying on closed-source models or large-scale labeled datasets, paving the way for automatic agent learning with open-source models. The framework demonstrates superior performance on complex question-answering tasks like HotpotQA and ScienceQA, with AUTOACT achieving significant improvements over other methods. The results indicate that moderate division-of-labor benefits group planning performance, and that AUTOACT's self-planning approach is more effective than prompt-based methods. The framework also shows that self-improvement and self-instruct can enhance knowledge and performance. AUTOACT is a promising approach for automatic agent learning, with potential for future research in expanding to more realistic task scenarios and improving synthetic trajectories through self-improvement.AUTOACT is an automatic agent learning framework for question-answering tasks that does not rely on large-scale annotated data or synthetic planning trajectories from closed-source models. It enables agents to learn from scratch using a limited set of user-provided data examples and a tool library. The framework starts with a META-AGENT that generates a large database through self-instruct, then synthesizes planning trajectories without external assistance. These trajectories are used to differentiate the META-AGENT into three sub-agents: PLAN-AGENT, TOOL-AGENT, and REFLECT-AGENT, each responsible for task decomposition, tool invocation, and self-reflection, respectively. The division-of-labor strategy allows for efficient and resource-effective learning. Comprehensive experiments with different LLMs show that AUTOACT performs better or on par with various strong baselines. Analysis confirms the effectiveness of the division-of-labor strategy, with AUTOACT's trajectory quality generally outperforming others. AUTOACT achieves self-planning without relying on closed-source models or large-scale labeled datasets, paving the way for automatic agent learning with open-source models. The framework demonstrates superior performance on complex question-answering tasks like HotpotQA and ScienceQA, with AUTOACT achieving significant improvements over other methods. The results indicate that moderate division-of-labor benefits group planning performance, and that AUTOACT's self-planning approach is more effective than prompt-based methods. The framework also shows that self-improvement and self-instruct can enhance knowledge and performance. AUTOACT is a promising approach for automatic agent learning, with potential for future research in expanding to more realistic task scenarios and improving synthetic trajectories through self-improvement.