| name | arena |
| description | Local multi-agent design or implementation competition. Use when several materially different solutions are plausible and selection quality matters. Creates isolated candidates, judges them against shared criteria, and lets the lead agent synthesize and verify the strongest result without cloud agents. |
Arena
Use competing local candidates only when diversity is worth the additional work.
Frame the arena
- Define the artifact being selected and four to six criteria tied to the real task.
- Give every candidate the same scope, repository evidence, constraints, and verification target.
- Select distinct approaches, not cosmetic variations.
Run candidates
- Use harness-native local subagents or locally installed model CLIs. Never claim a named model was
consulted unless it actually ran.
- Isolate implementation candidates in separate branches or worktrees. One writer owns each.
- Keep candidates bounded; they should return artifacts and evidence, not lengthy process diaries.
- Do not allow candidates to merge, push, deploy, or mutate shared external state unless the user
separately authorized it.
Judge and synthesize
When practical, have reviewers score candidates independently before seeing one another's verdicts.
The lead agent must still inspect all candidates, select the strongest base, graft only clearly better
elements, and run the repository's real verification. Report the criteria, winner, useful rejected
ideas, and evidence. If no candidate clears the bar, say so and redesign.
