The Architecture of Agent Harness Engineering: Execution Scaffolding for Autonomous Systems

Large language model (LLM) agents can now write and change real software, but how well they work depends on the code around the model, such as the loop that runs it, the tools it can use, the context it sees, and the checks on its work. Practitioners call this layer the agent harness. This paper turns recent research and practice into a simple six-part model of a harness and uses the Elephants and Goldfish principle as a design guide, in which project knowledge lives in files that the harness keeps and each working session gets a small, focused context. It then tests two harness ideas on 34 programming tasks from a public benchmark, using three LLMs with three trials each. Feeding unit test results back to the LLM raised the share of solved tasks from 63% to 94% on average, and from 34% to 80% for the weakest LLM. Asking the LLM to review its own code without running it gave no gain. A repair loop that kept only the latest code and test results did about as well as one that kept the full history and used about a third fewer input tokens. The results suggest that a harness should check work with real tests and keep each step's context small, instead of relying on the model to judge itself.