Imagine asking an AI assistant to plan a trip. It needs to not only search for flights and hotels but also use the results of each search to refine the next, like choosing a hotel near the chosen airport. This intricate dance of interconnected actions, where the output of one task becomes the input for another, is what we call nested API calls. Researchers are exploring this complexity with a new benchmark called NESTFUL. Current AI models, while impressive in handling single API calls, often stumble when faced with nested sequences. Think of it as a chain reaction – a small error in one step can snowball into a larger problem down the line. NESTFUL challenges Large Language Models (LLMs) with this realistic scenario using real-world APIs from areas like travel, finance, and social media, along with more abstract, non-executable API sequences. The benchmark tests the model's ability to select the right APIs, fill in the correct parameters, and execute them in the proper order, even when the required actions aren’t explicitly stated. Early results from NESTFUL reveal that even the most powerful LLMs struggle with this nested complexity. They often fumble with data types, misinterpret API specifications, and fail to chain the operations correctly. This highlights the need for better training data and refined algorithms that can manage these dependencies. However, there’s reason for optimism. NESTFUL provides a vital training ground for the next generation of AI models, allowing researchers to pinpoint weaknesses and design more sophisticated systems. As LLMs learn to navigate the intricacies of nested APIs, they’ll be better equipped to tackle the complex tasks we expect of truly intelligent assistants.