Harness 🤖 Agent Open source

What it does

Harness is a meta-skill for Claude Code that designs domain-specific agent teams. You describe a domain in plain language ("build a harness for this project") and it turns that description into a coordinated team of specialized agents plus the skills those agents use. Rather than being a fixed set of subagents, it is a factory that generates them: it analyses the domain, picks a team-architecture pattern, writes agent definitions into .claude/agents/, generates skills into .claude/skills/, and wires the orchestration between them.

What is inside

The plugin is a single skill with a six-phase workflow (domain analysis, team-architecture design, agent generation, skill generation, integration and orchestration, and validation) backed by reference files. It offers six architectural patterns: Pipeline, Fan-out/Fan-in, Expert Pool, Producer-Reviewer, Supervisor, and Hierarchical Delegation. Two execution modes are available: agent teams for work that needs inter-agent collaboration, and plain subagents for one-off tasks. Generated skills use progressive disclosure to keep context small, and the validation phase includes trigger verification, dry-run testing, and with-skill versus without-skill comparison. The README supplies ready-made prompts for domains such as deep research, full-stack websites, code review, documentation, and data-pipeline design.

Works with

Claude Code only, installed through its plugin marketplace. It requires Claude Code's experimental Agent Teams feature to be enabled.

How to install

/plugin marketplace add revfactory/harness
/plugin install harness@harness-marketplace

Maintenance and safety

Apache-2.0 licensed. The plugin writes agent and skill files into your project's .claude/ directory, which you can inspect before use. The author's effectiveness figures come from a small self-measured A/B study (n=15) with third-party replication still pending, and the README labels them as such.

Who should use it

Claude Code users who want to stand up a tailored multi-agent team for a new domain without hand-writing every agent and skill.