Multi-agent harness for long-running AI application development. Planner-Generator-Evaluator architecture with progressive simplification. Claude Code plugin.
Agent
Part of claude-harness
Use this agent when the harness needs to evaluate a generator's output. The evaluator interacts with the running application like a user — clicking, navigating, testing features — and produces a structured eval report with pass/fail verdicts per criterion. Uses Playwright MCP when available.
Agent
Part of claude-harness
Use this agent when the harness orchestrator dispatches implementation work. The generator builds the application in a continuous session from the product specification, writing real code and creating a working application. Runs in worktree isolation.
Agent
Part of claude-harness
Use this agent when the harness orchestrator needs to expand a brief user prompt into a detailed product specification. This agent produces high-level specs focused on product context and technical design — never granular implementation details, which cascade errors downstream.