Fan-out/fan-in distributes independent subtasks and later combines their results. It is effective for exploration, review, evaluation and option generation.
Pipelines are sequential: each agent transforms or checks the previous output. Pipelines are useful when roles have real dependencies.
The reviewer pattern has one agent create an artifact and another critique it against explicit criteria. The creator then repairs verified issues. Debate patterns use several agents to propose or challenge alternatives before an adjudicator decides; this may improve coverage but can amplify cost and correlated model bias.
