Arena 100 Claudes fight. One answer survives.

Same task, different minds, one bracket

A sub-agent is a copy of Claude that Claude Code starts to do one job and report back. Arena starts 100 of them by default and gives every one the exact same task, word for word. What makes them different is a strategy card: one of 15 reasoning modes, one of 12 workflows and one of 12 strategies. That's 2,160 possible cards, and no two agents get the same one.

1. 1

   Everyone answers

   Each agent writes its own answer, through its card. First principles, inversion, edge cases first, three drafts and pick one: 100 real attempts at your task.

2. 2

   They attack in pairs

   Answers are paired. Each side attacks the other: what is wrong, what requirement it missed, the exact case that breaks it. Up to 7 attacks, marked FATAL, MAJOR or MINOR.

3. 3

   They defend and fix

   Each side answers every attack, concedes what is right, and rewrites its answer to fix it. Conceding a real flaw scores better than defending it.

4. 4

   A judge scores both

   A separate judge reads both revised answers, checks every attack itself and scores them on a written rubric. It never sees the cards. The loser is out, and the survivors go to the next round.

   Correctness30

   Completeness25

   Robustness20

   Specificity15

   Clarity10

With 100 agents: 100 → 50 → 25 → 13 → 7 → 4 → 2 → 1. 7 rounds. An answer with a verified fatal flaw can't beat one without, whatever the scores say.

One message, or two commands

It needs Claude Code (the skill starts sub-agents, which the chat apps can't do) and Python 3.8 or newer. Nothing to sign up for, no API key, nothing to connect.

Three things before your first run

The skill is free. The tokens are yours.

Every answer, attack, defense and verdict is one sub-agent call that reads your task and one or two answers, and it all comes out of your Claude plan. Bigger tasks cost more per call. These are the exact numbers from bracket.py plan, which I ran for each size:

Add one call when there is an answer you rejected to beat. Claude Code runs at most 10 tool calls at once by default, so the arena runs in waves of 10, and a full 100 takes a while. The skill prints its size before it starts. When it starts by itself because you said “that's wrong”, it asks first: full arena, quick, or a normal retry.

For the answer that matters, after the retry failed

Six prompts for real work

Replace the part in brackets and paste into Claude Code. Each one already carries the rules the agents need, because they only see what you write.

What you get back

• The winning answer, in full.
• Why it won: the attacks it survived, its strategy card on one line, and the rounds it took.
• An honest comparison: if you started from an answer you rejected, a last judge compares the two blind and you see the scores, even when your old answer wins.
• The full record in .arena/<run>/: every answer, attack and verdict, if you want to look.

Honest limits

• It never touches your files. Every agent writes inside .arena/. Code changes come back as a diff in the winning answer, and the skill asks before applying anything.
• It's 100 copies of the model you're running, each with a different card. It doesn't call other AI models.
• It doesn't prove the answer is right. The judges are Claude too, scoring on a rubric you can read and change in the repository. You get the answer that survived every match, with the attacks it beat, so you can check.
• The same --seed gives the same cards and the same bracket. The answers can still differ.
• If your chat gets compacted halfway, nothing is lost: the whole tournament lives in one file and picks up where it stopped.

Short answers