Agent Arena
cebeuq/agent-arena
Arena de competição para agentes de código com verificação automática
Ver no GitHubHarness competitivo que executa múltiplos agentes de código (Claude, Codex, Cursor) em paralelo contra a mesma tarefa, em workspaces isolados com visibilidade do rival e condições de vitória verificadas. Ideal para benchmark de performance e otimização de agentes.
Instalação
npx convez-ai cebeuq/agent-arena1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Comparar performance de diferentes agentes IA em tarefas de codificação
- Otimizar benchmarks de código com validação automática
- Avaliar Claude, Codex e Cursor em condições iguais
Veredito

Ferramenta altamente especializada para developers que precisam benchmarkar agentes de código em ambiente controlado. Requer setup complexo (git, tmux, múltiplos CLIs) mas oferece insighs valiosos sobre performance relativa — útil para times que otimizam com IA.
README
A tradução para o português deste README ainda está na fila.
Agent Arena (WIP)
Agent Arena runs two coding-agent CLIs against the same task in isolated git worktrees. Each agent gets its own workspace, a live tmux pane, a refreshed read-only mirror of the rival workspace, and a claim command. The first agent to claim success and pass your verifier wins.
V1 includes built-in presets for:
- Claude Code:
claude - OpenAI Codex CLI:
codex - Cursor Agent CLI:
cursor-agent
Custom shell commands are also supported.
Install
npm install
npm run build
npm link
Agent Arena requires macOS or Linux plus git and tmux.
Quick Start
From the repository you want agents to work on:
agent-arena init
Edit arena.config.json:
{
"baseRepo": ".",
"baseRef": "HEAD",
"goal": "Make the benchmark at least 3x faster without changing output correctness.",
"verifyCommand": "npm test && npm run bench -- --min-speedup 3",
"agents": [
{ "id": "claude", "preset": "claude" },
{ "id": "codex", "preset": "codex" }
],
"peek": {
"refreshIntervalSeconds": 30
},
"tmux": {
"sessionPrefix": "agent-arena",
"attach": true
}
}
Start the match:
agent-arena start --config arena.config.json
Agent Arena will:
- Create two git worktrees under
.agent-arena/workspaces/<run-id>/. - Write each agent a
.arena/brief.md. - Start a tmux session with one pane per agent.
- Keep
.arena/rival/<rival-id>/refreshed from the rival workspace. - Accept the first passing claim as the winner.
Claiming A Win
Each workspace gets:
./.arena/claim.sh
Agents should run that command when they believe the goal is complete. The claim command runs verifyCommand inside the claiming agent's workspace. If it exits 0, that agent wins. If it fails, the claim is logged and the match continues.
Commands
agent-arena init
agent-arena agents list
agent-arena agents doctor --config arena.config.json
agent-arena start --config arena.config.json
agent-arena status --run <run-id>
agent-arena claim --run <run-id> --agent <agent-id>
status and claim can also take --state /path/to/state.json if the run is not in the global run index.
Command Templates
Agent commands can use these placeholders:
{goal}{goalFile}{claimCommand}{rivalDir}{workspace}{agentId}{runId}
Placeholders are shell-escaped automatically, so do not wrap them in extra quotes:
{
"id": "custom",
"command": "my-agent --prompt {goal} --cwd {workspace}"
}
Rival Mirrors
Agents do not receive writable access to the rival workspace. Instead, Agent Arena syncs the rival workspace into:
.arena/rival/<rival-id>/
The mirror is chmod'd read-only after each sync and excludes .git, .agent-arena, .arena, dependency folders, caches, and common secret files.
This is an accident-prevention boundary, not a hostile security sandbox. Processes running as the same OS user can usually change permissions back.
Examples
Example configs live in examples/:
claude-vs-codex.jsonclaude-vs-cursor.jsoncodex-vs-cursor.json
Notes
- The starting project must be a git repo.
- Worktrees are created from
baseRef; uncommitted changes in the source checkout are not copied. - V1 targets macOS and Linux only.
- Other agent CLIs are deferred, but custom commands work now.