Imagine asking an AI to write and run code to create a stunning visualization. Today's large language models (LLMs) would write the entire script before running it, like a chef meticulously preparing every ingredient before turning on the stove. But what if the AI could start cooking as soon as the first ingredient is ready? That's the idea behind Conveyor, a new system that makes LLMs dramatically faster when using external tools. Conveyor lets LLMs start using tools *before* they finish generating the full instructions. In our cooking analogy, this means the AI can start executing the first line of code while it's still writing the rest. This 'partial execution' significantly speeds up the entire process. We tested Conveyor with several tasks, like code generation, web searches, and complex planning. The results were impressive: Conveyor sped up some tasks by almost 40%! However, the benefits vary depending on the tool. For quick tasks like simple calculations or database lookups, the overhead of switching between tasks outweighs the gains from parallel execution. Conveyor's magic lies in its clever design. It provides a simple interface for tool developers to specify when partial execution is possible. Then, a smart scheduler orchestrates the process, making sure the LLM and the tools work together seamlessly without stepping on each other's toes. This is a big step forward in making LLMs more efficient and responsive. While Conveyor shines with certain tools, its effectiveness with others highlights an important area for future research: optimizing the interplay between LLMs and the ever-expanding toolkit they use to interact with our world.