Researchers have proposed using complete cyclic subtask graphs to improve the performance of large language model (LLM) agents in long-horizon tool-using tasks. These graphs allow for revisiting earlier subtasks and can be evaluated on various benchmark environments, including TextCraft, ALFWorld, and Finance-Agent. The study found that different tasks benefit from different workflow control strategies, and that complete cyclic subtask graphs are best used as a diagnostic tool to identify when backtracking is beneficial and when simpler controllers are more suitable. --- Why it matters: This research matters because it provides insights into the design of LLM agents for complex tasks, which can inform the development of more efficient and effective workflow control strategies. By understanding how different tasks benefit from various workflow control approaches, researchers and engineers can improve the performance of LLM agents in real-world applications. Source: https://arxiv.org/abs/2604.22820