Harness is a Claude Code plugin that functions as a team-architecture factory, automatically generating domain-specific agent teams and their associated skills based on natural language descriptions. When a user provides a domain description with prompts like "build a harness for this project," the plugin decomposes the request into coordinated teams of specialized agents and generates the necessary agent definitions and skills tailored to that domain. The plugin is written in HTML and operates at the L3 Meta-Factory layer of the Claude Code ecosystem, specifically within the Team-Architecture Factory sub-layer, distinguishing it from neighboring tools like Archon which focuses on runtime configuration.
The core functionality centers on six pre-defined team-architecture patterns that users can leverage: Pipeline for sequential dependent tasks, Fan-out/Fan-in for parallel independent tasks, Expert Pool for context-dependent selective invocation, Producer-Reviewer for generation followed by quality review, Supervisor for central agent with dynamic task distribution, and Hierarchical Delegation for top-down recursive delegation. The plugin automatically generates agent definitions in the .claude/agents/ directory and skills in the .claude/skills/ directory, implementing Progressive Disclosure for efficient context management. It includes validation features such as trigger verification, dry-run testing, and comparative testing between configurations with and without skills.
The repository is classified across multiple DevOps and infrastructure categories including feature flags, continuous delivery, deployment automation, cloud infrastructure, configuration management, release management, and CI/CD.
The plugin operates in two execution modes: Agent Teams mode as the default, which uses TeamCreate, SendMessage, and TaskCreate for scenarios requiring two or more collaborating agents, and Subagents mode for one-off tasks without inter-agent communication needs. The repository maintains version 1.2.0 under Apache 2.0 licensing and includes internationalization support with README documentation available in English, Korean, and Japanese.
Harness connects to a broader ecosystem of Claude Code tools and maintains relationships with neighboring repositories including coleam00/Archon, SaehwanPark/meta-harness, affaan-m/ECC, and wshobson/agents. The project has generated production applications including revfactory/harness-100, which contains 100 production-ready agent team harnesses across 10 domains in both English and Korean, totaling 200 packages with 1,808 markdown files. Research documented in revfactory/claude-code-harness demonstrates measured improvements in output quality, with reported metrics showing 60 percent average quality improvement, 100 percent win rate across 15 tasks, and 32 percent reduction in output variance, though these figures are explicitly noted as author-measured on a limited sample pending third-party replication.
