Do Competitively

This is a heavyweight quality-over-speed approach that spins up four parallel agents: three to generate competing solutions and a meta-judge to create custom evaluation rubrics. Each generator does self-critique loops before submitting. Then three judges evaluate all solutions using the rubrics, with verification steps to catch their own errors. The system adaptively decides whether to polish the winner, synthesize the best parts from multiple solutions, or start over if everything scored poorly. It claims 15-20% cost savings by skipping full synthesis when there's a clear winner. Use this when correctness matters more than time, like critical infrastructure code or high-stakes architecture decisions. The strictness of the orchestration is unusual: it literally threatens termination if you deviate from the process.