stanford-iris-lab/meta-harness
A framework for automated search over task-specific model harnesses that control what LLMs store, retrieve, and display during execution.

Meta-Harness provides a reusable framework for optimizing the scaffolding code that surrounds fixed base models in agent systems. It searches over harness configurations to find optimal combinations of memory, retrieval, and display mechanisms. The repository includes reference implementations for text classification harness optimization and Terminal-Bench 2.0 scaffold evolution, along with an onboarding process for applying the framework to new domains.