Reference code for the Meta-Harness paper.
Meta-Harness is a framework for automated search over task-specific model harnesses—the code that decides what information a fixed base model should store, retrieve, and use while working on a task. The framework is built in Python and designed to optimize harness code across different problem domains, from text classification to terminal-based command execution. The repository contains the reusable Meta-Harness framework along with two reference experiments from the associated paper by Lee, Nair, Zhang, Lee, Khattab, and Finn, with additional pilot work using the Harbor task framework. The codebase leverages language models like Claude and GPT variants as proposer agents to iteratively improve harness implementations, evaluated against task-specific benchmarks.