stanford-iris-lab/meta-harness

What it solves

Meta-Harness addresses the problem of optimizing the "harness"—the surrounding code that manages data retrieval, storage, and presentation—without needing to retrain or fine-tune the base AI model. Instead of changing the model itself, it focuses on improving the software scaffold that supports the model to maximize performance on specific tasks.

How it works

It provides a framework for automated search over task-specific model harnesses. It uses a proposer agent (such as Claude Code) to iteratively evolve and optimize the harness code. The process involves defining a domain specification and an evaluation plan, which the framework then uses to search for the most effective code structure to support the model's operations.

Who it’s for

Developers and researchers who want to improve the performance of a fixed base model on a specific domain or task without the overhead of training or fine-tuning.

Highlights

• Automated Search: Automatically searches for the best harness configuration for a given task.
• Reusable Framework: Includes an onboarding flow to help users apply the framework to new domains.
• Reference Experiments: Provides examples for text classification and Terminal-Bench 2.0 scaffold evolution.
• Extensible Proposer: Supports different proposer agents by adapting wrapper scripts.