Self-Evolving Agent Harness

Building agents that learn from experience is a central goal of artificial intelligence. Harness self-evolution offers a path toward this goal by adapting the system around a fixed base model. This article develops a unified framework for examining how execution feedback produces persistent harness changes. We organize algorithms by their update targets, distinguishing instructions, memory, and workflow specifications interpreted by fixed components from executable tools and runtime logic, and compare how they propose and admit changes. The theoretical analysis views harness evolution within the framework of evolutionary algorithms and draws on PAC learning theory for continual learning to provide insights into algorithm design. For evaluation, we distinguish four claims behind a reported gain: selection, whether the estimate is independent of the search or accounts for selection; activation, whether the accepted update operated in the scored runs; attribution, whether it caused the difference under controlled evaluation conditions; and improvement, whether the difference holds beyond the tasks that produced it. The reviewed protocols show why held-out performance, component use, and retained gains require different evidence. This article covers applications in personal assistance, enterprise automation, scientific discovery, robotics, and public services. We also discuss potential effects on productivity and skills, economic distribution and market structure, innovation and public services, privacy and security, and institutional governance. Drawing on these analyses, we identify future directions in update algorithms, supporting infrastructure, evaluation, and responsible deployment. The project repository is available at \url{https://github.com/zgcagi/Harness-RSI}.