nc://arena

Arena: a knockout tournament of sub-agents over one task — attack, defend and be judged until one answer wins. A Nodecode rewrite of J182Razor/hermes-arena-skill by J182Razor.

Get it

Tell the agent "add the arena cell", or run nodecode add nc://arena.

Use it

When an answer is not good enough, do not ask again: make it a tournament.

• plan says what a run costs and writes nothing. --quick is 16 competitors, about 91 sub-agent calls in 16 waves; the full arena is 100, about 595 calls. Just arena command=plan with no size uses the setting.
• init starts a run: it deals every competitor a different strategy card — a reasoning mode, a workflow and a strategy — and pairs round one.
• next says what to do now. For each phase it names, prompts writes the briefs and lists the jobs in waves; each job is one sub-agent call told to read its brief and follow it exactly. collect then reads the judges' verdicts, advance closes the round and pairs the survivors.
• Every round is attack, defend, judge. A judge scores both revised answers against the shipped rubric, and the cell does the arithmetic: the weighted total, the fatal rule, the tie-break. When one solution is left, winner names it, the card it won with, the attacks it survived and the matches it took — and, if there was an answer the user had already rejected, how the winner did against it, blind.
• The state is one file, arena.json, in the run folder under .arena/. A compacted conversation loses nothing: status and next resume the run.

The task file is what decides the result: sub-agents cannot see the conversation, so init's task must stand alone — the request in the user's own words, every constraint, the context a stranger needs, and what "done" looks like.

Set it up

Nothing is required. The arena section of the config refines it: agents (the default size, 100) and wave (sub-agent jobs launched at once, 10).

What is in it

License

MIT. The original is Copyright (c) 2026 Jake Schincariol: see LICENSE and NOTICE.