meta-harness

Why we included this project

Meta-Harness is a research framework for automating the search over task-specific model harnesses, the code wrapped around a fixed base model that controls what it stores, retrieves, and shows while working. Instead of hand-tuning that scaffolding, it runs an automated loop that proposes and evaluates candidate harnesses for a given task. The repo ships the reusable framework plus two worked reference experiments, one searching memory systems for text classification and one evolving scaffolds for Terminal-Bench 2.0, so you can see the loop in action before adapting it. It is research code rather than a polished product, and the authors only verify that it runs, so expect to invest in setup. If you build LLM agents and are tired of hand-editing the scaffolding around them, this is a useful starting point and a clear reference for the underlying paper.