Researchers have developed a method called Evo-Harness that allows self-improving large language model agents to learn from experience in complex real-world tasks. The approach involves updating a structured harness across sequential tasks, which enables the agent to distill noisy and single-shot executions into reusable skill harnesses for adaptation across domains and topics. This formulation allows for a systematic study of key self-improvement factors through the proposed Evo-Harness method. --- Why it matters: This matters because it provides a principled understanding of how large language model agents can effectively learn on the fly in complex real-world tasks, which is crucial for developing capable and self-improving AI systems. Source: https://arxiv.org/abs/2608.15071