| name | Best of N (parallel attempts) |
| description | Delegate N parallel subagents on the same task, then pick the best result. |
| triggers | ["best of n","parallel attempts","explore options","try variants","best-of-n"] |
Best of N (parallel attempts)
Use this when a task has a wide solution space and you want the strongest result: run N
independent attempts in parallel and choose the best. Each attempt is isolated, so they
cannot interfere with each other.
Steps
- Pick N. Default N = 3 (cap to the parallelism limit). For a code task use the
implementer role; for analysis/design use researcher or design-doc-writer.
- Fan out.
delegate N subagents with the same task in a single delegate
call so they run in parallel. Give each identical, complete instructions.
- Compare. When the summaries return, evaluate each against the goal and the
acceptance criteria (correctness, simplicity, completeness, risk).
- Choose. Pick the best attempt. The others' isolated copies are simply rejected at
merge time (not applied).
- Summarize. State which attempt you chose and why, and what is staged for merge.
Notes
- The attempts are blind to each other — that is the point (diversity).
- Vary the instruction slightly per attempt only if you want different angles; otherwise
keep them identical for a fair comparison.
- Only the chosen attempt's changes should be merged; explicitly discard the rest.