Automated search for the code that wraps your LLM

What it does

Meta-Harness is a framework that automatically searches for task-specific model harnesses — the code surrounding a fixed base model that decides what to store, retrieve, and show while the model works. It provides a reusable search loop and two reference experiments from the paper: one that searches memory systems for text classification, and another that evolves scaffolds for Terminal-Bench 2.0. The optimized Terminal-Bench 2 harness from the paper is maintained in a separate artifact repository.

The interesting bit

Instead of optimizing prompts or fine-tuning weights, Meta-Harness treats the entire surrounding infrastructure as the variable to evolve. The onboarding workflow even instructs you to hand ONBOARDING.md to your coding assistant and let it draft a domain_spec.md — a refreshingly candid automation of the setup process.

Key highlights

• Searches over harness code for a frozen base model, not weights or prompts
• Reference experiments cover memory-system search and terminal-task scaffold evolution
• Includes an ONBOARDING.md designed to be fed to a coding assistant for domain adaptation
• Proposer agent interactions are wrapped and logged; examples currently assume Claude Code
• Optimized paper artifacts are split into a separate repo to keep the core framework focused

Caveats

• The authors state this is a cleaned-up paper release that has only been verified to run, not thoroughly tested
• Reference examples assume Claude Code as the proposer agent; other agents require a new wrapper
• The optimized Terminal-Bench 2 harness is not in this repo; it lives in a separate artifact repository

Verdict

Useful if you are building agent infrastructure and want to automate the search for better scaffolding around a fixed model. Less relevant if you need battle-tested production tooling or are only looking to tweak prompts.