Imagine teaching a robot to perform complex tasks, not through meticulous programming or constant hand-holding, but by offering occasional, helpful hints. That's the intriguing idea behind new research from Tencent exploring how to train AI agents using 'weak supervision' from large language models (LLMs). Traditionally, training AI agents for real-world scenarios has been a resource-intensive endeavor, demanding either expert demonstrations or precise feedback signals. This new approach takes a different tack. Instead of relying on perfect examples or explicit rewards, the agents learn by interacting with their environment and receiving feedback from a 'critic' LLM. This critic observes the agent's actions and selects the most promising attempts, guiding the agent towards better performance over multiple iterations. Think of it as a coach offering constructive criticism rather than prescribing exact moves. The researchers tested their method on a dataset involving thousands of APIs for various applications, from web search to weather forecasting. Impressively, their agents, trained with minimal supervision, achieved performance close to that of GPT-4, despite using significantly smaller language models. This suggests that weak supervision could be a powerful tool for developing more adaptable and efficient AI agents. While the method shows promise, challenges remain. The iterative training process can be computationally expensive, and the critic's feedback, while helpful, isn’t always perfect. Future research could explore refining the critic's judgment and improving the efficiency of the training process. This innovative training method represents a step toward building AI agents that can learn and adapt more like humans do, opening doors to a wide range of applications where traditional training methods fall short.