Harness v2
A community Claude Code plugin that analyzes a project and generates specialized agents, skills, and orchestration rules across workflow, persistent-agent, and one-shot subagent modes.
Project overview
Harness v2 is a team-architecture generator for Claude Code. It inspects existing agents, skills, and CLAUDE.md entries, then chooses among fixed workflows, persistent collaboration, or one-shot subagents before generating specialist agents, supporting skills, and an orchestrator. It is best suited to projects with recurring complex work and maintainers who are prepared to own project-level agent configuration. Start on a test branch, review every planned file change, and run a small team before expanding the setup.
Core capabilities
Project-level team design
Harness inspects existing agents, skills, and CLAUDE.md before deciding whether to create, extend, or maintain a configuration and split the project into specialist roles.
Three execution modes
Fixed lists and validation flows can use Workflow, long-running collaboration can use named agents and SendMessage, and one-result tasks can use parallel subagents.
Team patterns and quality checks
The documentation defines six team patterns and requires failure handling, structural validation, trigger tests, realistic request tests, and at least one failure scenario.
Feedback-driven evolution
The evolve skill reads harness files, git history, and intermediate artifacts, turns recurring feedback into reusable rules, and applies changes one at a time with verification.
v1 configuration migration
The migration guide replaces TeamCreate, TeamDelete, and experimental flags with Workflow, Agent, SendMessage, and shared tasks.
Installation and usage
Install Harness v2 from the Claude Code Marketplace in a test project, then ask it to design a small, bounded team. Review the generated files, routing rules, and execution mode before expanding the configuration into a primary project.
Let an AI Agent install it
Send this prompt to Codex, Claude Code, or another AI agent that can work with your local environment.
Help me install Harness v2 from https://github.com/revfactory/harness. First read README.md, docs/quickstart.md, .claude-plugin/plugin.json, and skills/harness/SKILL.md. Check the current Claude Code version and existing .claude configuration, then use the repository's documented Marketplace commands. Run /plugin list afterward and generate only a minimal harness in a test project, listing every added or modified file. Ask me before requesting credentials, additional permissions, Workflow execution, overwriting files, or launching a large parallel agent run.- Latest Claude Code
- Access to github.com
- A test project where .claude/agents, .claude/skills, and CLAUDE.md may be created or modified
- Enough token and call quota for a small multi-agent verification run
/plugin marketplace add revfactory/harness
/plugin install harness@harness-marketplace- 1Add the Marketplace and install the plugin
Run /plugin marketplace add revfactory/harness and then /plugin install harness@harness-marketplace in Claude Code. Run /plugin list and confirm that harness appears.
- 2Inspect existing project configuration
Open the test project and check .claude/agents, .claude/skills, and CLAUDE.md. Record files that must not be overwritten and routing rules that must remain.
- 3Generate a minimal team
Give Claude Code a bounded request, such as creating an implementation-and-review harness for the current project. Ask it to show the roles, execution mode, and target paths before writing files.
- 4Review and run a small verification
Check that agents, skills, the orchestrator, and CLAUDE.md agree and that no TeamCreate, TeamDelete, or experimental flags remain. Run one small task and inspect intermediate artifacts, failure handling, and the final result.
- 5Evolve one issue at a time
After a real run, invoke the evolve skill. Change one class of issue at a time and retain the rationale and test result so one-off feedback does not become an overfitted rule.
Run /plugin list and confirm that harness appears. In a test project, ask Claude Code to “build a harness for this project,” then check for .claude/agents, .claude/skills, and a CLAUDE.md entry. Review the diff before running any large team.
- The README does not document an uninstall command, so this page does not infer one.
- Workflow requires explicit user consent, and generated configuration should be checked for conflicts before it is written.
- Multi-agent tests may consume tens or hundreds of thousands of tokens. Start with a two- or three-role flow.
Use cases
Separate implementation and review
Create producer and verifier roles so one agent proposes a change, another reviews it independently, and the orchestrator reconciles the results.
Batch research and synthesis
Distribute a known list of sources across agents, then pass structured results or files into a synthesis phase.
Persistent specialist collaboration
Use named agents for architecture or content work that requires multiple rounds of feedback, direct messages, and shared task state.
Migrate a legacy harness
Find v1 TeamCreate, TeamDelete, and experimental flags, then reconnect agents and skills through the three v2 execution modes.
Assessment
This assessment is based on the README, Quickstart, both skills, the migration guide, and the CHANGELOG; it does not present documentation review as hands-on use. Harness v2's strongest feature is that team design, skill authoring, orchestration, testing, and later iteration live in one project-level system with explicit checks for duplicate roles and legacy configuration. The tradeoffs are equally visible. Generation touches project configuration, multi-agent runs consume more tokens, and results still depend on the host model, granted tools, and maintainers reviewing the generated files.
Why it may be useful
- The installation path is short, while manifests, core skills, migration guidance, and change history remain publicly auditable
- Execution-mode and team-pattern choices have concrete criteria, with clear output paths and verification checklists
- The evolve skill provides a complete path from runtime feedback to incremental changes and regression checks
What to know first
- Capabilities were verified from repository documentation only; installation and end-to-end behavior were not tested
- Generated artifacts modify project configuration, leaving overwrite handling, permissions, and version control to the adopter
- The author's small v1 comparison cannot establish consistent v2 gains across different codebases
README
Harness v2
Overview
A community Claude Code plugin that analyzes a project and generates specialized agents, skills, and orchestration rules across workflow, persistent-agent, and one-shot subagent modes. A meta-skill that designs domain-specific agent teams, defines specialized agents, and generates the skills they use.
Getting started
- Install Harness v2 from the Claude Code Marketplace in a test project, then ask it to design a small, bounded team. Review the generated files, routing rules, and execution mode before expanding the configuration into a primary project.
- Harness v2 designs domain-specific agent teams and stores roles, working methods, and routing rules in project-level configuration.
- It supports workflow orchestration, persistent-agent collaboration, and one-shot subagents, with mixed modes across phases.
- Run /plugin list and confirm that harness appears. In a test project, ask Claude Code to “build a harness for this project,” then check for .claude/agents, .claude/skills, and a CLAUDE.md entry. Review the diff before running any large team.
Configuration
/plugin marketplace add revfactory/harness
/plugin install harness@harness-marketplace