LLM agent systems usually select actions from a fixed set, limiting their capabilities in real-world scenarios. This paper introduces a framework that allows dynamic creation and composition of actions by generating and executing programs in a general-purpose language at each step. Results on the GAIA benchmark show improved flexibility and performance, enabling agents to adapt to unforeseen situations. The framework outperforms previous methods, allowing agents to recover when predefined actions fail. The authors currently hold the top spot on the GAIA leaderboard, showcasing the effectiveness of their approach. To access the code, visit the provided URL.
https://arxiv.org/abs/2411.01747