Imagine an AI agent capable of tackling complex tasks, from navigating mazes to solving intricate math problems. TaskGen, a novel agentic framework, makes this a reality by breaking down large tasks into smaller, manageable subtasks. Unlike other agentic systems that rely on free-form text output, TaskGen leverages StrictJSON, a concise JSON format, to enhance communication efficiency and reduce processing overhead. This innovative approach not only streamlines the workflow but also minimizes token usage and improves accuracy. At its core, TaskGen employs a hierarchical structure of agents. A 'meta agent' delegates subtasks to specialized 'inner agents' or 'equipped functions,' each designed for specific tasks. This division of labor mirrors a well-organized team, where each member contributes their expertise to achieve a common goal. The framework also incorporates a shared memory system, allowing agents to access and modify information on a need-to-know basis. This prevents information overload and improves decision-making. TaskGen's memory bank further strengthens its capabilities, providing a repository of information across various levels of abstraction. This allows the agent to dynamically access relevant information, enhancing its ability to learn and adapt to new situations. Furthermore, Global Context provides the agent with real-time updates on the environment, enabling it to adjust its actions accordingly. This adaptability is especially valuable in scenarios like navigating a dynamic maze, where obstacles can change unexpectedly. TaskGen has been rigorously tested across a diverse range of environments, demonstrating its versatility and effectiveness. From achieving a 100% solve rate in dynamic maze navigation to tackling complex math problems with 71% accuracy, TaskGen consistently outperforms other AI agents. It also excels in text-based game environments and demonstrates promising results in web browsing and question answering tasks. The future of TaskGen is bright, with plans for integrating better planning capabilities, multi-modal support, and fostering collaboration between multiple agents with diverse skills. This collaborative approach promises to unlock new levels of AI performance and problem-solving. Join the growing TaskGen community on GitHub and Discord to contribute to its development and explore the future of task-based AI.