Compound Engineering
EveryInc/compound-engineering-pluginEvery's plugin of 36 skills for a brainstorm, plan, build, review and capture-the-learning loop, with an optional hands-off /lfg pipeline.
Recommended
Broad access Depth 4/5
Harness Claude skills
A Claude Code meta-skill that designs a project-specific agent team, writes the agent definitions and skills for it, and has an evolve skill to refine them from feedback.
Our verdict Recommended with caveats
A substantial and well-documented generator of agent teams, but it is Claude Code specific, its skill text is in Korean, and the repo has no evals to show the generated harnesses beat a plain setup.
What Harness does
Harness is a plugin with two skills. The main one, harness, is triggered by requests like "build a harness for this project" (the trigger phrases are mostly Korean, and the README is available in English and Korean). It walks through eight phases: audit any existing harness, analyze the domain, choose an execution mode and team architecture, write agent definitions into .claude/agents/, generate skills into .claude/skills/, add orchestration plus a pointer in CLAUDE.md, run verification, and later maintain the result.
The design step picks among six team patterns (pipeline, fan-out/fan-in, expert pool, producer-reviewer, supervisor, hierarchical delegation) and three execution modes: Workflow scripts for deterministic fan-outs and verification loops, persistent named agents that exchange messages, or one-shot sub-agents. It also assigns a model tier per agent instead of pinning everything to one model. Generated harnesses are meant to include quality patterns such as adversarial verification, judge panels and loop-until-dry sweeps, and verification can compare a skill with and without it.
The second skill, evolve, reads the generated agents, skills and git history, collects your feedback on results, generalizes it and edits the harness, keeping a change log and avoiding reverting earlier fixes. The harness SKILL.md is large (about 35 KB) with nine reference files; the repo has no evals or CI. Version 2 removed the dependency on the old experimental agent-teams flag and includes a migration path from v1.
Good for
- Claude Code users who want a custom multi-agent setup for a recurring domain task
- Teams that want agent roles and skills generated and then tuned from usage feedback
- People comfortable reading Korean-language skill text or prompting in Korean or English
Not for
- One-off tasks where a single prompt or an existing skill is enough
- Users of agents other than Claude Code, since it relies on Claude Code's agent and Workflow features
What's inside depth 3/5
Two skills, one of them about 35 KB with nine reference files, but no scripts, tests, evals or CI.
- Skills
- 2
- Reference docs
- 9
- Scripts
- 0
- Tests / evals
- 0
- Templates
- 0
- Avg SKILL.md
- 20.7 KB
Counted from the repository file tree. How we score depth.
Dependencies and costs
- Claude Code with sub-agent and Workflow support
Security review
No issues foundWe scanned all 29 files in this repository and read every script. Nothing in it reaches beyond what the skill describes.
A markdown-only meta-skill (plus a plugin manifest) that designs agent teams and generates agent and skill files under the project's .claude/agents and .claude/skills, and records a change log in the project's CLAUDE.md. It has no scripts, no allowed-tools pre-approvals, no hooks and no network use. Generated files are written into your project, so review them before committing.
What we checked
The repository (29 files) on Oct 29, 2026. We had every file and read every script.
NVIDIA SkillSpector static scan: 2 pattern hits (MCP Rug Pull 1, Agent Snooping 1). The two hits are an 'npx markdownlint' line in the PR template and an 'ls -la .claude/agents/ .claude/skills' command in the quickstart docs; neither is executed by the skill.
A review, not a certification. Skills change after we check them — read the files before installing.
License
Apache-2.0 — commercial use allowed.
Can it run unattended?
Yes. The Workflow execution mode it generates supports verification loops, budgets, structured outputs and resume support, though how unattended the result is depends on what it generates for your domain.
Alternatives to Harness
task-observer (One Skill to Rule Them All)
rebelytics/one-skill-to-rule-them-allA meta-skill that watches your sessions, logs corrections and repeated work, and proposes improvements to your other skills for you to approve.
Recommended
Check first Depth 4/5
wshobson/agents
wshobson/agentsA multi-harness plugin marketplace with 94 plugins, 202 agents, 185 skills and 105 commands for Claude Code, Codex, Cursor, OpenCode, Copilot, Antigravity and Pi.
With caveats
Check first Depth 2/5
Superpowers
obra/superpowersA 15-skill development methodology that makes coding agents clarify the goal, write a plan, work through it with subagents and verify before calling it done.
Recommended
Check first Depth 3/5
FAQ
What is Harness for Claude Code?
Harness is a meta-skill plugin that turns a description of your project or domain into a team of specialized agents plus the skills they use. It writes them into .claude/agents and .claude/skills and adds orchestration instructions.
How do I install Harness?
Run /plugin marketplace add revfactory/harness and then /plugin install harness@harness-marketplace. You can instead copy skills/harness and skills/evolve into ~/.claude/skills; no environment variables or experimental flags are needed.
Is Harness safe to install?
It is markdown-only with no scripts, so there is no hidden code to run. The risk is in what it generates, because agent definitions can carry tool permissions that you should read before use.
Harness vs Compound Engineering: which should I use?
Harness generates a custom agent team and skills for a given domain and then evolves them from your feedback. Compound Engineering ships a fixed set of workflow commands and agents, so it suits you if you want a ready process rather than a generator.
Sources: GitHub repository (stars, license, file tree, last push Sep 28, 2026). Review based on the README and SKILL.md files as published.