arena is for when Claude keeps giving a weak answer. Instead of asking again and again, it gives the same task to N sub-agents, each with a different strategy, and makes them fight in a bracket until one solution is left.
Invoke
Terminal
$arena <task>
$arena --quick <task>
What it does
- Spins up N sub-agents (100 by default, 16 with
--quick, any number with--agents N), each with its own strategy card: a reasoning mode, a workflow, and a strategy. - Runs a single-elimination bracket: solutions attack each other, defend and revise, and a judge scores every match on a written rubric.
- Keeps every piece of bookkeeping in a stdlib
bracket.pythat only writes to.arena/, so a run can be resumed or replayed with--seed. - Triggers: “arena”, an answer you are not satisfied with, “try again”, “do better”, or asking agents to compete.
Based on Jakeschincariol/arena-skill (MIT).
Install
Included in the free bundle:
Terminal
npx aiblueprint-cli@latest agents setup
Install only this skill:
Terminal
npx skills@latest add Melvynx/aiblueprint --skill arena