Voting and Consensus

On genuinely ambiguous calls — design choices, naming, when to refactor — running three agents independently and taking the majority answer is more reliable than asking one. The cost is 3× the tokens for that decision; the benefit is variance reduction. Use it sparingly, on calls where being wrong is expensive.

Don't use voting for execution. Three builders writing the same code is wasteful, not robust. Voting belongs at decision points: "should we extract this helper?" "is this name better than that?" "is this migration reversible?"