Harness

Harness is a Claude Code plugin that acts as a team-architecture factory: given a one-sentence description such as 'build a harness for this project,' it turns a domain description into a purpose-built agent team and the skills those agents use.

Version 2 targets three native execution modes that Claude Code ships: deterministic workflow orchestration scripts for fan-outs and verification loops, persistent named-agent collaboration with shared task lists and retained context across turns, and lightweight one-shot sub-agent delegation. It codifies workflow quality patterns such as adversarial verification, judge panels, loop-until-dry and completeness critics into generated harnesses, assigns each agent a model tier based on the task's complexity and latency needs instead of pinning every agent to the same model, and can detect and migrate an existing v1 harness.

Core features include six agent-team architecture patterns such as pipeline, fan-out/fan-in, expert pool and hierarchical delegation, skill generation using progressive disclosure, data-passing and error-handling protocols, verification through evals and A/B testing, and an evolve command that captures the difference between a harness's initial and current state and feeds that feedback back into its agents and skills.