What problem does it solve? Unit tests and code review cannot tell you whether a fresh agent can actually use your CLI, library, or agent harness. This Skill closes that gap by running disposable subagents against a measurable scenario, treating their friction as the specification, and iterating until the tool's signal quality converges. ## Core Features & Use Cases - Closed-loop subagent validation: Spawn fresh agents with explicit read/forbidden file lists, a round budget, and a deterministic success criterion, then compare convergence across tool versions. - Friction-driven fixing: File one issue per agent complaint, commit per friction point with the agent's verbatim quote, and write versioned validation reports under docs/reports/. - Convergence tracking and stopping criteria: Track per-agent metrics in a comparison table, tighten round budgets as the tool matures, and stop when diminishing returns appear. - Use Case: You built a CLI whose output agents must parse. Run this loop: agent-a reports the notation is ambiguous, you fix it in one commit, agent-b confirms convergence improved from 10.3% to 0.2% diff. ## Quick Start Ask the agent to run the agent-validation-loop on your tool by defining a fixture with a brief and goal artifacts, then spawning a fresh subagent to attempt the task.
