Harness Forge is a Claude Code skill that runs an end-to-end harness-optimization loop — propose → score → keep the Pareto-best → repeat — to improve the code around a fixed model: its memory, retrieval, context construction, summarization, prompt templates, and tool-selection logic. The model never changes; the scaffolding gets better.

It is a native reimplementation of the method in Meta-Harness: End-to-End Optimization of Model Harnesses (Lee, Nair, Zhang, Lee, Khattab & Finn, 2026). The original reference repo ships ~1,260 lines of Python (claude_wrapper.py + meta_harness.py) whose job is to drive a headless Claude: spawn a session, parse its output, track tool calls, log everything, loop. Inside Claude Code, that runtime already exists as first-class tools. So Harness Forge keeps only the irreducible domain logic — a cheap scorer — and expresses the entire outer loop as native orchestration. The whole search becomes ~75 lines instead of ~1,260.

The idea in one picture

seed the frontier with the incumbent harness (the thing to beat)
repeat:
    PROPOSE   k candidate harness variants     ← parallel proposer agents write code
    VALIDATE  each imports / type-checks
    SCORE     each on a held-out-protected eval ← a $0, deterministic scorer
    FRONTIER  Pareto-merge: quality up, cost down, floor-respecting
final: score the frontier once on the untouched test split

The proposer is the mutation operator. The frontier is the search memory. The model is frozen throughout — which is exactly why this fits a fixed / off-the-shelf-API deployment, where you can't change the weights and the gain has to come from the harness.

The paper's headline result was +7.7 accuracy points at ~4× fewer context tokens on text classification — a pure harness-side win. Harness Forge reproduces that shape of result natively.

Why native?

claude_wrapper.py is a hand-rolled agent runtime. Claude Code is an agent runtime. So every orchestration piece has a native equivalent, and the Python driver becomes redundant:

The only thing you still write is the cheap scorer + rubric + candidate interface. Everything orchestration-shaped is free.

Quick start

1. Install the skill — one line:

curl -fsSL https://raw.githubusercontent.com/001TMF/harness-forge/main/install.sh | bash

Or as a Claude Code plugin (inside Claude Code):

/plugin marketplace add 001TMF/harness-forge
/plugin install harness-forge@tmf-skills

# project-scoped (./.claude/skills, this repo only)
curl -fsSL https://raw.githubusercontent.com/001TMF/harness-forge/main/install.sh | bash -s -- --project

# via skills.sh (vercel-labs/skills)
npx skills add 001TMF/harness-forge --skill meta-harness -a claude-code

# manual
git clone https://github.com/001TMF/harness-forge.git
cp -r harness-forge/skills/meta-harness ~/.claude/skills/meta-harness

It auto-triggers when you talk about optimizing a harness, scaffold, prompt system, memory or retrieval policy, or summarizer — or invoke it directly as the meta-harness skill.

2. Run the worked example ($0, no model, no network):

cd harness-forge/examples/memory-summary
python score_baselines.py
# -> baseline_incumbent  fidelity=1.000 chars=269   (the system to beat)

3. Run a real search — invoke the Workflow tool with the example's loop script:

Workflow({ scriptPath: "<abs>/examples/memory-summary/native_meta_harness_workflow.js",
           args: { dir: "<abs>/examples/memory-summary", rounds: 2, k: 3 } })

Proposer agents run on your Claude subscription; the scorer is $0; there is no solver model and no metered API. A successful round produces a compressor holding fidelity at < 269 chars.

What you supply (the five blocks)

The loop is native; the domain is yours. Templates are in skills/meta-harness/assets/; how-to is in references/building-blocks.md:

1. Candidate interface — one clean, swappable boundary (an ABC / Protocol).
2. A $0 deterministic scorer + rubric — the inner loop; runs hundreds of times, so no LLM, no network. It must vary with the candidate (see the trap below).
3. An eval corpus with a held-out split.
4. A proposer prior — a mini-skill steering proposers toward mechanism-level changes (not constant-tuning) and forbidding eval-set leakage.
5. A frontier + run log — the state. scripts/pareto.py computes the floor-respecting frontier deterministically.

The one trap that sinks naive harness searches

The frozen-replay defect. If your scorer replays cached outputs (a recorded run, a frozen trace), a scaffolding candidate cannot change the recorded result — only the cost axis moves. A naive "maximize quality, minimize cost" search then wins by emptying the context while the frozen quality score never drops, producing a confident, meaningless frontier.

Test: "If I swap in a wildly different candidate, can this number change for a quality reason?" If only cost can move, you are replaying frozen outputs.

Fix: grade something the candidate genuinely controls (retrieval relevance, compression fidelity, a counterfactual decision), and/or run quality as a one-sided do-no-harm floor rather than a maximize axis. The skill makes this — plus held-out discipline, an anti-Goodhart floor, and anti-leakage — load-bearing. Full treatment in references/method.md.

Repository layout

harness-forge/
├── .claude-plugin/marketplace.json   # installable as a Claude Code plugin
├── install.sh                        # one-line curl|bash install
├── skills/
│   └── meta-harness/             # the installable skill
│       ├── SKILL.md              #   what/when, the loop, the 5 blocks, the guardrails
│       ├── references/           #   method · native-execution · building-blocks · worked example
│       ├── assets/               #   templates: workflow loop, scorer, interface, proposer prior
│       └── scripts/pareto.py     #   reusable floor-respecting Pareto frontier
└── examples/
    └── memory-summary/           # a complete, runnable search (the $0 demo + the native loop)

When to use this (and when not)

Use it when the base model is fixed, there are repeated tasks, and a cheap measurable eval exists (or can be built) — i.e. the gain has to come from the harness. Classic targets: context bloat, weak retrieval, lossy summarization, brittle prompt scaffolds.

Don't when the gain must come from the model weights (do RL / fine-tuning instead), or when there is no stable evaluation loop. Meta-Harness and RL are complementary: in a fixed-base-model phase, Harness Forge is the only available optimizer — and it forces the eval-hardening a later RL phase also depends on, at near-zero cost. See references/method.md §6.

Credit

The method is Meta-Harness by Yoonho Lee, Roshen Nair, Qizheng Zhang, Kangwook Lee, Omar Khattab, and Chelsea Finn. This repo is an independent native reimplementation as a Claude Code skill; it vendors no code from the original repo. If you use it, please cite the paper:

@misc{lee2026metaharnessendtoendoptimizationmodel,
  title={Meta-Harness: End-to-End Optimization of Model Harnesses},
  author={Yoonho Lee and Roshen Nair and Qizheng Zhang and Kangwook Lee and Omar Khattab and Chelsea Finn},
  year={2026},
  eprint={2603.28052},
  archivePrefix={arXiv},
  primaryClass={cs.AI},
  url={https://arxiv.org/abs/2603.28052},
}

• Paper: https://arxiv.org/abs/2603.28052
• Reference implementation: https://github.com/stanford-iris-lab/meta-harness

License

MIT © 2026 Tristan Farmer