Harness is a meta-skill designed as a Claude Code plugin that facilitates the creation of domain-specific agent teams by defining specialized agents and generating the skills they utilize. Its primary purpose is to leverage Claude Code's agent team system to decompose complex tasks into coordinated teams of specialized agents. When prompted with a request like "build a harness for this project," it automatically generates agent definitions (stored in .claude/agents/) and skill definitions (in .claude/skills/) tailored to the specified domain.

The core methodology of Harness is structured around a 6-phase workflow:

1. Phase 1: Domain Analysis: This initial phase involves analyzing the problem domain to understand its specific requirements, constraints, and objectives.
2. Phase 2: Team Architecture Design: In this phase, Harness designs the overall architecture for the agent team, deciding between collaborative "Agent Teams" (for complex tasks requiring inter-agent communication) and "Subagents" (for simpler, one-off tasks without inter-agent communication). This involves selecting appropriate architectural patterns.
3. Phase 3: Agent Definition Generation (.claude/agents/): Harness proceeds to generate detailed definitions for each specialized agent within the team. These definitions specify the agent's role, responsibilities, and capabilities, saved as Markdown files in the .claude/agents/ directory (e.g., analyst.md, builder.md).
4. Phase 4: Skill Generation (.claude/skills/): Following agent definition, Harness generates the specific skills that each agent will use. These skills are designed with Progressive Disclosure, a technique to manage context efficiently by revealing information incrementally as needed, and are saved in the .claude/skills/ directory (e.g., analyze/skill.md, build/skill.md).
5. Phase 5: Integration & Orchestration: This phase focuses on integrating the defined agents and their skills, establishing inter-agent data passing mechanisms, error handling protocols, and team coordination protocols to ensure seamless execution of the overall task.
6. Phase 6: Validation & Testing: The final phase involves validating the generated harness. This includes trigger verification, dry-run testing, and comparative tests (with-skill vs. without-skill execution) to ensure functionality and efficacy.

Harness offers several key features:

• Agent Team Design: It supports 6 architectural patterns for agent collaboration:
  • Pipeline: For sequential, dependent tasks.
  • Fan-out/Fan-in: For parallel, independent tasks.
  • Expert Pool: For context-dependent selective invocation of agents.
  • Producer-Reviewer: Involves a generation phase followed by a quality review.
  • Supervisor: A central agent dynamically distributes tasks.
  • Hierarchical Delegation: Supports top-down recursive delegation of tasks.
• Skill Generation: Automatically generates skills, employing Progressive Disclosure for optimized context management.
• Orchestration: Manages inter-agent data flow, error handling, and defines coordination protocols.
• Validation: Provides tools for trigger verification, dry-run testing, and comparative performance analysis (with and without the generated skills).

Harness operates in two primary execution modes:

• Agent Teams (default): Recommended for two or more agents requiring collaboration, utilizing TeamCreate, SendMessage, and TaskCreate functionalities.
• Subagents: Suitable for one-off tasks that do not require inter-agent communication, involving direct agent tool invocation.

Output files generated by Harness are structured within a project's .claude/ directory, containing agent definition files in .claude/agents/ and skill files in .claude/skills/.

Research conducted on Harness demonstrates its effectiveness in enhancing LLM code agent output quality through structured pre-configuration. An A/B testing experiment across 15 software engineering tasks showed significant improvements:

• Average Quality Score: Increased by 60% (from 49.5 to 79.3).
• Win Rate: Achieved 100% (15/15 tasks).
• Output Variance: Decreased by 32%.

The research indicates that the effectiveness scales with task complexity, with greater improvements observed for advanced and expert-level tasks ($+29.6\%$ and $+36.2\%$ respectively, compared to $+23.8\%$ for basic tasks).

Installation can be done via the Claude Code marketplace (/plugin marketplace add revfactory/harness then /plugin install harness@harness) or directly by copying the skills/harness directory to ~/.claude/skills/harness/. It requires $CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1$ to be enabled and is licensed under Apache 2.0. The plugin structure includes a plugin.json manifest, skill definitions in SKILL.md, and reference documentation on agent design patterns, orchestrator templates, team examples, and skill authoring/testing guides.