A task is a single bounded transformation that one evaluated model call performs reliably. What should an expert consider before adding multiple agents?

1. Remove every deterministic validation step
2. Whether the additional coordination has a justified benefit over the simpler design
3. Assume parallel agents always guarantee a correct consensus
4. Add agents solely to maximize their number