AI agents can now find their way through thousands of tools

Researchers built a new way to test AI agents that use many software tools, and a new method to help those agents navigate complex tasks. This means AI systems can now complete multi-step jobs more reliably, even when they have thousands of tools to choose from.

AI models are good at using individual tools, but they struggle when they need to string together many steps using a vast library of options. This paper gives researchers a way to measure that struggle and a new algorithm to make AI agents better at it. It means future AI systems could handle much more complex, multi-step tasks in areas like e-commerce or scientific discovery.