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 language agents to learn from scratch by synthesizing planning trajectories without human or strong closed-source model assistance. The framework employs a division-of-labor strategy to differentiate into sub-agents with distinct functions: task decomposition, tool invocation, and self-reflection. This approach allows for efficient and resource-conscious learning, as demonstrated by comprehensive experiments with various large language models (LLMs), where AUTOACT outperforms or matches strong baselines in performance. The division-of-labor strategy is effective, with AUTOACT's trajectory quality generally outperforming others. AUTOACT's META-AGENT starts with self-instruct to generate a database of task examples, then synthesizes planning trajectories and differentiates into sub-agents. The framework's ability to self-plan without external data or models makes it suitable for real-world scenarios where data is limited. Experiments on HotpotQA and ScienceQA show that AUTOACT achieves better performance than existing methods, with the quality of trajectories generated by AUTOACT being comparable to those from GPT-4. The framework's approach to division-of-labor and self-improvement is validated through extensive experiments and human evaluations, demonstrating its effectiveness in enhancing agent performance and efficiency. AUTOACT's design allows for flexible task handling and efficient learning, making it a promising solution for automatic agent learning in complex question-answering tasks.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 language agents to learn from scratch by synthesizing planning trajectories without human or strong closed-source model assistance. The framework employs a division-of-labor strategy to differentiate into sub-agents with distinct functions: task decomposition, tool invocation, and self-reflection. This approach allows for efficient and resource-conscious learning, as demonstrated by comprehensive experiments with various large language models (LLMs), where AUTOACT outperforms or matches strong baselines in performance. The division-of-labor strategy is effective, with AUTOACT's trajectory quality generally outperforming others. AUTOACT's META-AGENT starts with self-instruct to generate a database of task examples, then synthesizes planning trajectories and differentiates into sub-agents. The framework's ability to self-plan without external data or models makes it suitable for real-world scenarios where data is limited. Experiments on HotpotQA and ScienceQA show that AUTOACT achieves better performance than existing methods, with the quality of trajectories generated by AUTOACT being comparable to those from GPT-4. The framework's approach to division-of-labor and self-improvement is validated through extensive experiments and human evaluations, demonstrating its effectiveness in enhancing agent performance and efficiency. AUTOACT's design allows for flexible task handling and efficient learning, making it a promising solution for automatic agent learning in complex question-answering tasks.