Harness is a team-architecture factory for Claude Code. Say "build a harness for this project" (English) or "하네스 구성해줘" (한국어) or "ハーネスを構成して" (日本語), and the plugin turns your domain description into an agent team and the skills they use — picked from six pre-defined team-architecture patterns.
Harness leverages Claude Code's agent team system to decompose complex tasks into coordinated teams of specialized agents. Say "build a harness for this project" and it automatically generates agent definitions (.claude/agents/) and skills (.claude/skills/) tailored to your domain.
Harness lives at the L3 Meta-Factory layer of the Claude Code ecosystem — the layer that generates other harnesses rather than being one. Inside L3, we pick a specific sub-layer: Team-Architecture Factory.
| Layer | What it does | Neighbors we coexist with |
|---|---|---|
| L3 — Meta-Factory / Team-Architecture Factory (us) | Domain sentence → agent team + skills, via 6 pre-defined team patterns | — |
| L3 — Meta-Factory / Runtime-Configuration Factory | Deterministic, repeatable runtime configurations | coleam00/Archon |
| L3 — Meta-Factory / Codex Runtime Port | Same concept, Codex runtime | SaehwanPark/meta-harness |
| L2 — Cross-Harness Workflow | Standardize skills/rules/hooks across multiple harnesses | affaan-m/ECC |
Archon generates deterministic runtime configurations. Harness generates team architectures (pipeline, fan-out/fan-in, expert pool, producer-reviewer, supervisor, hierarchical delegation) plus the skills agents use. Different sub-layers of the same L3. Pick Archon for runtime determinism, Harness for team architecture, or combine them.
- Agent Team Design — 6 architectural patterns: Pipeline, Fan-out/Fan-in, Expert Pool, Producer-Reviewer, Supervisor, and Hierarchical Delegation
- Skill Generation — Auto-generates skills with Progressive Disclosure for efficient context management
- Orchestration — Inter-agent data passing, error handling, and team coordination protocols
- Validation — Trigger verification, dry-run testing, and with-skill vs without-skill comparison tests
The harness evolution mechanism feeds deltas (what worked / what didn't) back into the factory, so the next generation is measurably better. When a generated harness is used in a real project, the /harness:evolve skill captures the delta between the initial architecture and the shipped one, and feeds it back into the factory so the next generation for a similar domain starts closer to the shipped state.
Initial harness ──▶ Real project use ──▶ Shipped harness
│
▼ (delta capture via /harness:evolve)
┌───────────────┐
│ Factory │◀── better next-gen draft
└───────────────┘
We call this the Harness Evolution Mechanism (KR: 하네스 진화 메커니즘; JA: ハーネス進化メカニズム).
Phase 1: Domain Analysis
↓
Phase 2: Team Architecture Design (Agent Teams vs Subagents)
↓
Phase 3: Agent Definition Generation (.claude/agents/)
↓
Phase 4: Skill Generation (.claude/skills/)
↓
Phase 5: Integration & Orchestration
↓
Phase 6: Validation & Testing
A. Upstream (just the harness skill)
/plugin marketplace add revfactory/harnessB. Fork — includes the Korean Persona Injection branch (recommended until PR #9 lands)
/plugin marketplace add hongsw/harnessThe fork's main is fast-forwarded to feat/korean-persona-injection, so no branch ref is needed. This adds korean-persona-search / korean-voice-adapter / korean-persona-harness (3 skills), Codex CLI compatibility, and verification artifacts on top of the existing harness. After PR #9 merges into upstream, switch back to A (revfactory/harness). Details: Korean Persona Injection section below.
⚠️ Already installedrevfactory/harness? Uninstall it first, then install the fork. Both marketplaces use the same name (harness-marketplace) and plugin name (harness), so adding the fork on top would conflict or shadow the upstream. Clean migration in 4 steps:/plugin uninstall harness@harness # 1. Remove the plugin /plugin marketplace remove harness-marketplace # 2. Remove the marketplace /plugin marketplace add hongsw/harness # 3. Add the fork's marketplace /plugin install harness@harness # 4. Reinstall (now from fork)To switch back after PR #9 merges: repeat steps 1–2, then
add revfactory/harness. The dataset cache (~/.cache/korean-persona-search/) persists, so no redownload needed.
/plugin install harness@harness# Copy the skills directory to ~/.claude/skills/harness/
cp -r skills/harness ~/.claude/skills/harnessharness/
├── .claude-plugin/
│ └── plugin.json # Plugin manifest
├── skills/
│ └── harness/
│ ├── SKILL.md # Main skill definition (6-Phase workflow)
│ └── references/
│ ├── agent-design-patterns.md # 6 architectural patterns
│ ├── orchestrator-template.md # Team/subagent orchestrator templates
│ ├── team-examples.md # 5 real-world team configurations
│ ├── skill-writing-guide.md # Skill authoring guide
│ ├── skill-testing-guide.md # Testing & evaluation methodology
│ └── qa-agent-guide.md # QA agent integration guide
└── README.md
Trigger in Claude Code with prompts like:
Build a harness for this project
Design an agent team for this domain
Set up a harness
| Mode | Description | Recommended For |
|---|---|---|
| Agent Teams (default) | TeamCreate + SendMessage + TaskCreate | 2+ agents requiring collaboration |
| Subagents | Direct Agent tool invocation | One-off tasks, no inter-agent communication needed |
| Pattern | Description |
|---|---|
| Pipeline | Sequential dependent tasks |
| Fan-out/Fan-in | Parallel independent tasks |
| Expert Pool | Context-dependent selective invocation |
| Producer-Reviewer | Generation followed by quality review |
| Supervisor | Central agent with dynamic task distribution |
| Hierarchical Delegation | Top-down recursive delegation |
Files generated by Harness:
your-project/
├── .claude/
│ ├── agents/ # Agent definition files
│ │ ├── analyst.md
│ │ ├── builder.md
│ │ └── qa.md
│ └── skills/ # Skill files
│ ├── analyze/
│ │ └── SKILL.md
│ └── build/
│ ├── SKILL.md
│ └── references/
Copy any prompt below into Claude Code after installing Harness:
Deep Research
Build a harness for deep research. I need an agent team that can investigate
any topic from multiple angles — web search, academic sources, community
sentiment — then cross-validate findings and produce a comprehensive report.
Website Development
Build a harness for full-stack website development. The team should handle
design, frontend (React/Next.js), backend (API), and QA testing in a
coordinated pipeline from wireframe to deployment.
Webtoon / Comic Production
Build a harness for webtoon episode production. I need agents for story
writing, character design prompts, panel layout planning, and dialogue
editing. They should review each other's work for style consistency.
YouTube Content Planning
Build a harness for YouTube content creation. The team should research
trending topics, write scripts, optimize titles/tags for SEO, and plan
thumbnail concepts — all coordinated by a supervisor agent.
Code Review & Refactoring
Build a harness for comprehensive code review. I want parallel agents
checking architecture, security vulnerabilities, performance bottlenecks,
and code style — then merging all findings into a single report.
Technical Documentation
Build a harness that generates API documentation from this codebase.
Agents should analyze endpoints, write descriptions, generate usage
examples, and review for completeness.
Data Pipeline Design
Build a harness for designing data pipelines. I need agents for schema
design, ETL logic, data validation rules, and monitoring setup that
delegate sub-tasks hierarchically.
Marketing Campaign
Build a harness for marketing campaign creation. The team should research
the target market, write ad copy, design visual concepts, and set up
A/B test plans with iterative quality review.
Harness is not alone in the Claude Code / agent-framework ecosystem. The following repos live in adjacent layers; each is described in a parallel "X is …, Harness is …" form so you can pick the one that fits your need or combine several.
| Repo | Their position | Relationship to Harness |
|---|---|---|
| coleam00/Archon | "harness builder" — deterministic, repeatable runtime configurations | Same L3, neighbor sub-layer. Archon is a Runtime-Configuration Factory, Harness is a Team-Architecture Factory. Pick Archon for runtime determinism, Harness for team architecture, or combine them. |
| SaehwanPark/meta-harness | Codex port of the same concept | Same L3, different runtime. Use Harness on Claude Code, meta-harness on Codex. |
| affaan-m/ECC | "Agent harness performance & workflow layer" (sits on top of existing harnesses) | Different layer. ECC is a standardization layer across harnesses; Harness is a factory that generates harnesses. Serial combination possible. |
| wshobson/agents | Subagent / skill catalog (182 agents, 149 skills) | Factory ↔ parts supply. wshobson is a catalog to shop from; Harness designs the team. Absorb wshobson entries as parts inside a Harness-generated team. |
| LangGraph | State-graph orchestration, LLM-agnostic | Different track. LangGraph is for long-running, state-recoverable orchestration; Harness is for fast Claude-Code-native team design. |
revfactory/harness-100 — 100 production-ready agent team harnesses across 10 domains, available in both English and Korean (200 packages total). Each harness ships with 4-5 specialist agents, an orchestrator skill, and domain-specific skills — all generated by this plugin. 1,808 markdown files covering content creation, software development, data/AI, business strategy, education, legal, health, and more.
revfactory/claude-code-harness — A controlled experiment across 15 software engineering tasks measuring the impact of structured pre-configuration on LLM code agent output quality.
| Metric | Without Harness | With Harness | Improvement |
|---|---|---|---|
| Average Quality Score | 49.5 | 79.3 | +60% |
| Win Rate | — | — | 100% (15/15) |
| Output Variance | — | — | -32% |
Key finding: effectiveness scales with task complexity — the harder the task, the greater the improvement (+23.8 Basic, +29.6 Advanced, +36.2 Expert).
Exact phrasing to use everywhere: +60% avg quality (49.5 → 79.3), 15/15 win-rate, −32% variance (n=15, author-measured A/B, third-party replications pending).
Full paper: Hwang, M. (2026). Harness: Structured Pre-Configuration for Enhancing LLM Code Agent Output Quality.
The fork hongsw/harness:feat/korean-persona-injection adds 3 non-invasive skills that inject NVIDIA Nemotron-Personas-Korea (1M rows, CC BY 4.0) into harness-generated agents at runtime — yielding agent definitions whose voice, workplace manner, generational vocabulary and cultural cues actually read as Korean.
A blind comparison of the same 5-person standup-meeting task (102 vs 103 lines, same LLM, same workload) showed the persona-injected team produced richer voice differentiation (5 distinguishable speakers vs 5 indistinct), 4 inter-personal exchanges vs 0 (mentoring, gratitude, family-aware on-call negotiation), and Korean-specific manners (단정 회피, 컨펌 톤, 우회 표현) that the generic team did not produce. See _workspace/comparison_test/01_team_output_comparison.md.
# Clone the fork branch and run the one-liner installer
git clone -b feat/korean-persona-injection https://github.com/hongsw/harness.git
cd harness
./scripts/install-korean-persona.sh --target both # Claude Code + Codex CLI
# OR install directly via Codex's skill-installer
python3 ~/.codex/skills/.system/skill-installer/scripts/install-skill-from-github.py \
--repo hongsw/harness \
--path skills/korean-persona-search \
--path skills/korean-voice-adapter \
--path skills/korean-persona-harness
# OR add the fork as a Claude Code plugin marketplace
# /plugin marketplace add hongsw/harness
# /plugin install harness
# Dataset cache (~few GB, shared between both runtimes; first run only)
pip install huggingface_hub pyarrow
python3 $SKILL_DIR/korean-persona-search/scripts/download.pyOnce installed, both harnesses coexist. Pick by intent (or invoke explicitly when ambiguous):
| Task | Skill | Trigger phrase |
|---|---|---|
| Generic / English / language-neutral team | harness (existing) |
"build a harness for this project" |
| Korean users / Korean-market scenarios / persona simulation | korean-persona-harness (new) |
"한국어 페르소나로 팀 만들어줘" / "build a Korean persona team" |
| Mixed (English skeleton + a few Korean personas) | harness for structure → call korean-persona-search for select roles |
invoke skills sequentially |
The two skills are auto-routed by description keywords — Korean-persona terms (한국어 페르소나, 한국 시나리오, Nemotron Korea) trigger the new skill; everything else falls back to the original harness. State the language explicitly when in doubt.
Compatible with both Claude Code and Codex CLI (same SKILL.md format). Details in docs/korean-persona-injection.md and docs/install-korean-persona.md. The Korean README (README_KO.md) has the full breakdown including the comparison table and verbatim Korean dialogue examples.
- Agent Teams enabled:
CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1
Q1. Isn't "+60%" oversold?
A. The +60% figure comes from an author-measured A/B (n=15, 15 tasks, measured on the sister repo claude-code-harness). Every citation in this repo pairs the number with the disclosure "n=15, author-measured, third-party replications pending" in the same sentence. For adoption decisions, we recommend running a 2–4 week internal pilot and measuring your own numbers.
Evidence:
- Author A/B: revfactory/claude-code-harness
- Paper: Hwang, M. (2026). Harness: Structured Pre-Configuration for Enhancing LLM Code Agent Output Quality
Q2. Why "harness factory" and not "harness builder"? Isn't this competing with Archon?
A. Archon generates deterministic runtime configurations — it's a Runtime-Configuration Factory. Harness generates agent team architectures (team structure, message protocols, review gates) — it's a Team-Architecture Factory. They are neighbor sub-layers of the same L3 Meta-Factory and serve different needs. Pick Archon for runtime determinism, Harness for team-architecture patterns, or combine them (design architecture with Harness → deploy runtime with Archon).
Evidence:
- Archon self-definition: clawfit docs/reference-levels.md
- Sub-layer declaration: see the Category — Where Harness Sits section above
- Archon repo: github.com/coleam00/Archon
Q3. Isn't "Claude Code only" too narrow? What about Gemini/Codex?
A. Currently the official runtime is Claude Code only. A Codex port of the same concept — SaehwanPark/meta-harness — is already public, so Codex teams can start there. Harness chose "Claude-Code-native, deep" over "multi-runtime, shallow"; cross-runtime collaboration with sibling repos (meta-harness, harness-init, OpenRig) is on the roadmap.
Evidence:
- Codex port: github.com/SaehwanPark/meta-harness
- Cross-runtime scaffolder: github.com/Gizele1/harness-init
Apache 2.0