Harness is a meta-skill that designs domain-specific agent teams, defines specialized agents, and generates the skills they use — all from a single prompt.
From domain analysis to validation, Harness automates the entire lifecycle of building multi-agent systems.
6 battle-tested patterns — Pipeline, Fan-out/Fan-in, Expert Pool, Producer-Reviewer, Supervisor, and Hierarchical Delegation.
Auto-generates skills with Progressive Disclosure, keeping context lean and loading details only when needed.
Message-based, task-based, and file-based data passing between agents with built-in error handling and fallback strategies.
Trigger verification, dry-run testing, and with-skill vs without-skill A/B comparison tests to prove quality gains.
Copy any prompt into Claude Code after installing Harness. Click to expand.
Multi-agent investigation teams
Full-stack build pipeline
Creative team for episodic content
End-to-end video planning
Parallel multi-dimension audit
Docs from codebase analysis
End-to-end data workflows
Campaign asset creation & testing
Say “build a harness” and Harness runs a structured 6-phase pipeline to produce production-ready agent teams.
Analyze your codebase, identify task types, detect user expertise, and prevent conflicts with existing agents.
Choose between Agent Teams and Subagents, select from 6 architectural patterns, and define agent separation criteria.
Generate .claude/agents/*.md files with roles, principles, I/O protocols, and team communication contracts.
Create skills with YAML frontmatter, pushy trigger descriptions, and Progressive Disclosure (metadata → body → references).
Wire agents together with orchestrator skills, data-passing protocols, error handling, and team size guidelines.
Structure checks, trigger verification, dry-run tests, with/without-skill comparison, and iterative refinement.
Pick the right coordination strategy for your workload. Click a pattern to learn more.
Sequential dependent tasks
Parallel independent tasks
Context-dependent selection
Generate then verify
Central dynamic dispatch
Top-down recursive split
Each agent's output feeds directly into the next. Ideal for workflows where steps are strictly ordered — code generation → review → testing → deployment. Simple to reason about, but the chain is only as fast as the slowest link.
A controlled A/B experiment across 15 software engineering tasks at 3 difficulty levels.
100% win rate across all 15 test cases
32% less variance — more consistent, predictable output
Scales with complexity — the harder the task, the bigger the gain
Production-ready agent teams across 10 domains. 200 packages (EN + KO). 1,808 files. All generated by this plugin.
Two commands. That’s it.
Then just say: “Build a harness for this project”