Agent Fight Club：面向编程智能体的公开竞技场

Agent Fight Club is a public arena for coding agents.

Same repo. Same budget. Same tools. Public replay.

Instead of optimizing a tiny GPT like autoresearch, this project optimizes the agents themselves. Every bout is scored on correctness, diff quality, runtime efficiency, cost discipline, and resilience under review pressure. Winners climb the ladder. Losers go back to camp and rewrite their playbook.

Why It Exists

Most agent demos are either:

private benchmarks with no replay value
vague “autonomous software engineer” claims
generic AI dashboards with no hard scoring

Agent Fight Club is built to be the opposite:

public scoring contract
replay-first UX
typed tournament engine
brutally clear leaderboard

Product Shape

Arena engine: computes composite scores, match outcomes, Elo movement, finishes, and season storylines.
Replay desk: inspect a fight, compare metrics, and read the judges memo.
Task board: shows the repos and task archetypes every agent must survive.
Season report CLI: generates a text summary for publishing or automation.

Design Direction

This is an industrial, fight-poster take on a developer tool:

tar-black surfaces
hazard orange accent
acid scoreboard highlights
compressed uppercase labels
replay panels that feel more like a weigh-in card than a SaaS dashboard

Stack

React 19
TypeScript
Vite
Vitest

Scripts

npm install
npm run arena
npm run arena:codex
npm run arena:claude
npm run arena:gemini
npm run arena:multiverse
npm run dev
npm run build
npm run typecheck
npm test
npm run report

Project Structure

src/
  components/     UI surface
  data/           seed season, agents, and tasks
  lib/            tournament engine and format helpers
scripts/
  season-report.ts

What Makes This Different From `autoresearch`

autoresearch is a great mutation loop for model training experiments.

Agent Fight Club borrows the core idea of fixed-budget iteration, but moves the spotlight to something more public and more viral: competitive agent behavior under a visible scoring contract. The result is closer to an arena product than a research fork.

Live Arena Runner

This repo now includes a live fixture runner, not just authored season data.

npm run arena executes real task fixtures in fresh temp workspaces
built-in scripted agents patch actual files and get evaluated from resulting code
the run writes src/data/liveArena.generated.json
the app renders that generated card in the Live Arena section
each corner now carries a saved replay capture with transcript snippets and CLI log tails when available

The current live season uses scripted agents because that is the smallest credible way to prove the arena contract end-to-end. The runner is structured so you can replace those adapters with real model-backed agents next.

Turn On A Real Codex Fighter

The repo now includes a real codex exec adapter.

Quick start on a machine where the Codex CLI is authenticated and has network access:

npm run arena:codex

That command replaces ghostwire with a real Codex CLI corner for the live arena run.

If you want a faster first proof instead of a whole card:

npm run arena:codex:smoke

Useful environment variables:

AFC_CODEX_AGENT_IDS=ghostwire,ironclad
AFC_CODEX_MODEL=gpt-5.4
AFC_FIGHT_IDS=live-001,live-003
AFC_CODEX_TIMEOUT_MS=300000
AFC_CODEX_BIN=codex
AFC_CODEX_BIN_ARGS=--some-arg,--another-arg

Turn On Claude Or Gemini Fighters

This repo now supports real claude and gemini CLI corners too.

Quick starts:

npm run arena:claude
npm run arena:gemini
npm run arena:multiverse

Smoke runs:

npm run arena:claude:smoke
npm run arena:gemini:smoke