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

Makes self-evolving agent harnesses auditable: automatically links each harness behavior to its implementing source via static analysis + LLM annotation, so an evolved harness stays readable, navigable, and editable by humans rather than accreting opaque generated scaffolding. Tencent HY LLM Frontier with Indiana University, Maryland, UGA, and NUS. Part of the July 2026 harness-evolution cluster (see the AI2/UW evaluation critique, arXiv 2607.12227).