Annie Wang argues that an AI agent can improve at its job without a new model, by changing its harness: the instructions, context, tools, and software around the model. The article was published November 13, 2026 on O'Reilly Radar, and it is a reflective essay rather than a news report of an event.
Wang says a self-evolving agent needs three things. The first is evaluation, to tell whether a change helped. Wang uses a restaurant-planning exercise to make the point. An agent suggests a dinner place and time for a group, and Wang says a recommendation that met everyone's needs was marked wrong because it differed from a sample answer. Wang argues for checking requirements instead, such as price, dietary needs, and timing. The second is an improvement process that can do more of the work on its own, such as reviewing past tasks, proposing changes, and testing them. The third is the set of harness parts the agent may change.
Wang describes ModularRSI, which divides the harness into five parts: context management, the agent loop, tool use, observation management, and task completion detection. Wang says these parts can be studied separately before being combined. Wang also describes "dreaming," a process in which a system reviews and organizes saved memories between sessions. The article names Harness-RSI and Anthropic