Researchers introduce Dive, a framework that trains AI agents using diverse and verifiable tasks, significantly improving how large language models interact with external tools and complex workflows.