Researchers have developed a new framework called FlowEvo that allows large language model agents to adapt and learn from complex tasks without needing extensive training. This is achieved through the co-evolution of workflows and executable skills at inference time. Successful workflows are compiled into reusable skills, which can be stored and used in future tasks. The system also tracks the utility of each skill and suppresses those that cause negative transfer. --- Why it matters: This matters to engineers and researchers because it enables more efficient and adaptive learning in complex environments, potentially leading to improved performance and reduced computational resources required for large language models. Source: https://arxiv.org/abs/2607.21596