# Meta-Harness

Reference code for the Meta-Harness paper.

Repository: https://github.com/stanford-iris-lab/meta-harness
Canonical: https://ross.abutalabs.com/products/meta-harness
Homepage: https://yoonholee.com/meta-harness/
Language: Python
License: MIT
License Family: permissive
Topics: harness-engineering, llm-agents
Last push: 2026-07-11T22:25:06+00:00
Link (homepage): https://yoonholee.com/meta-harness/

## Health v2 (maintenance only)
Score: 56/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 92, release rhythm 35, longevity 10
- inputs: {"age_days": 140, "days_push": 53, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases, young
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1464, forks 140 (observed 2026-08-28T04:04:47.864678+00:00)

## What it is
Meta-Harness is a Python framework for automated end-to-end search over task-specific model harnesses — the code surrounding a fixed base LLM that decides what to store, retrieve, and show while the model works. It includes an onboarding flow for adapting the search loop to new domains and two reference experiments from the accompanying paper: memory-system search for text classification and scaffold evolution for Terminal-Bench 2.0.

## Use cases
- automatically optimize an LLM agent scaffold for a benchmark
- search over memory-system designs for text classification
- evolve terminal-bench agent harnesses to improve task scores
- apply harness optimization to a new domain with an onboarding flow
- diagnose agent failures from raw execution traces and propose fixes
- reproduce results from the Meta-Harness paper

## When to choose
- you have a fixed base model and want to automatically improve the surrounding harness code
- you want to reproduce or extend the Meta-Harness paper experiments
- you need an LLM-driven search loop that reasons over full execution traces rather than compressed summaries
- you want a starting point for harness engineering on benchmarks like Terminal-Bench 2.0

## When to avoid
- you need a production-ready agent framework with polished APIs and support guarantees
- you want to fine-tune model weights rather than optimize harness code
- your project cannot use Claude Code or requires a different proposer agent without adaptation work
- you need a lightweight library with minimal dependencies — the search loop is compute- and evaluation-intensive

## Facets
- artifact type: framework
- maturity: active
- function: agent-framework, llm-inference, machine-learning, benchmarking, developer-tools
- domain: large-language-models, machine-learning, developer-tools
- platform: python, cli
- tags: harness-engineering, llm-agents, automated-search, scaffold-optimization, terminal-bench, research-code, evolutionary-search, ai-agents, automation, linux, macos

## Member repositories
- stanford-iris-lab/meta-harness (main) score 56
- stanford-iris-lab/meta-harness-tbench2-artifact (examples) score 48

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:04:47.864678+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T04:35:07.970562+00:00, confidence not recorded.
  - readme: https://github.com/stanford-iris-lab/meta-harness (fetched 2026-08-28T04:04:47.864678+00:00, sha d20509ed3e21)
  - homepage: https://yoonholee.com/meta-harness/ (fetched 2026-08-29T11:43:29.096369+00:00, sha 729264d079b4)
- Data as of 2026-08-30T08:39:29.467469+00:00.