Harness

Why it's ranked here

Harness offers a concrete design process instead of a loose prompt collection. Its pattern catalog covers sequential, parallel, selective, reviewed, supervised, and hierarchical work. The generated artifacts remain editable Markdown. The main reservation is its dependence on an experimental Claude Code feature for true team execution.

What's good

The pattern choices map cleanly to common work shapes, which helps users reason about coordination before generating anything. Harness also addresses data passing, errors, review gates, skill context management, and validation. It can use simpler subagents for isolated work, while collaborative tasks get shared team coordination.

Tradeoffs

Full team behavior requires an experimental environment flag. Without it, generated teams fall back to single-agent execution and several coordination patterns silently break. Multi-agent runs may fan out across several model calls and consume substantial tokens. The official runtime targets Claude Code, not a portable or model-neutral execution layer.

How to use it well

Use Harness when a recurring, complex Claude Code workflow needs explicit roles, handoffs, and review stages. Start with a narrow domain sentence, inspect the generated Markdown, and test one realistic task before broader adoption. Choose subagents for one-off work. Look elsewhere for deterministic runtime configuration, state recovery, or cross-runtime orchestration.

Observed

License

Apache License 2.0

Packaging

Claude Code plugin, also installable by copying its global skill directory

Interface

Claude Code CLI prompts and plugin marketplace commands

Runtime support

Claude Code v2.x or later; official runtime support is Claude Code only

Artifact format

Generates local Markdown agent definitions, skills, references, and orchestration configuration

Execution modes

Agent Teams, direct subagents, and documented hybrid combinations