Harness Learning: Agent Harnesses Are Learnable Policy Components

Agent performance depends on its harness, the execution system surrounding the model that organizes context, tools, memory, and control flow. Recent methods aim to improve subsequent execution through program search, configuration adaptation, editor training, or accumulated experience, but differ in modified components, update mechanisms, and reuse evaluation. We argue that harnesses are learnable policy components and formulate Harness Learning as a common learning problem with three complementary dimensions: Representation specifies executable components and available modifications; Optimization uses experience to improve harnesses or their construction mechanisms; Generalization evaluates resulting systems on new tasks. A shared execution policy and task score connect these dimensions. The framework analyzes existing systems and motivates research on harness composition, transfer, continual improvement, and model–harness co-learning.