Harness Handbook: Making Evolving Agent Harnesses Readable, Navigable, and Editable

This paper argues that an agent’s capability depends not only on its foundation model, but also on the harness that builds prompts, manages state, invokes tools, and coordinates execution. It introduces the Harness Handbook, a behavior-centric representation generated through static analysis and LLM-assisted structuring, and Behavior-Guided Progressive Disclosure (BGPD), which navigates agents from high-level behaviors to implementation details while checking locations against current source. On modification requests involving two open-source harnesses, the authors report better behavior localization and edit-plan quality with fewer planner tokens, especially for scattered code, rarely executed paths, and cross-module interactions.