agentarena

agentarena

Race your AI agents. Any agent, any task, your data.

$ agentarena run

 agentarena v0.1.0 — racing 3 agents on 3 tasks
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

 Task 1/3: fix-type-error
   claude-code ····· PASS   15s   $0.08   4.2K tokens   2 calls
   aider ··········· PASS   23s   $0.14   8.7K tokens   5 calls
   codex ··········· PASS   31s   $0.21  12.1K tokens   8 calls

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 RESULTS
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

              Pass Rate   Avg Time   Avg Cost   Total Tokens
 claude-code  3/3 100%      45s       $0.20       34.6K
 aider        2/3  67%      69s       $0.35       65.2K
 codex        2/3  67%      71s       $0.42       82.2K

 Winner: claude-code (highest pass rate, lowest cost)

Why agentarena?

Every company building AI is asking the same question: which agent actually works best?

Today that answer is opinions, blog posts, and vibes. Your manager asks for a POC — you spend two weeks manually testing three tools and write a Google Doc that says "I think Claude was better."

agentarena gives you hard numbers in 30 minutes:

Inspired by the ActionEngine paper which found 11.8x cost differences and 5.67x token usage variance between agent architectures on identical tasks. agentarena makes these differences visible on your own data.

Install

pip install agentarena

Quick Start

1. Create a bench.yaml in your project:

agentarena init

2. Define your tasks and agents:

project: my-app
timeout: 120

tasks:
  - name: fix-type-error
    prompt: "Fix the TypeScript type error in src/auth/login.ts"
    validate: "npx tsc --noEmit"

  - name: add-pagination
    prompt: "Add offset/limit pagination to GET /api/users endpoint"
    validate: "bun test test/api/users.test.ts"

agents:
  - name: claude-code
    command: "claude --print --max-turns 10 '{prompt}'"
    patterns:                                          # optional: extract metrics
      tokens_in: "input tokens:\\s*([\\d,]+)"
      tokens_out: "output tokens:\\s*([\\d,]+)"
      cost: "cost:\\s*\\$?([\\d.]+)"

  - name: aider
    command: "aider --message '{prompt}' --yes-always --no-git"

  - name: my-custom-agent                              # any CLI tool works
    command: "my-tool run '{prompt}'"

3. Run the race:

agentarena run

How It Works

For each task x agent combination:

1. Creates a clean sandbox (git worktree for code repos, temp directory for anything else)
2. Runs the agent CLI with your prompt
3. Runs your validation command (tests, typecheck, lint — anything with an exit code)
4. Collects metrics: wall time, tokens, cost, LLM calls, pass/fail
5. Cleans up the sandbox

Works with any project — git repos, plain directories, any language, any domain.

CLI

agentarena run                          # Race all agents on all tasks
agentarena run --task fix-type-error    # Run specific task
agentarena run --agent claude-code      # Run specific agent
agentarena run --json                   # Export as JSON
agentarena run --csv                    # Export as CSV
agentarena run --md                     # Export as Markdown
agentarena init                         # Create starter bench.yaml
agentarena history                      # List past runs

Metric Extraction

agentarena uses regex patterns defined in your config to extract metrics from agent output. No code changes needed for new agents:

agents:
  - name: my-agent
    command: "my-agent '{prompt}'"
    patterns:
      tokens_in: "Input:\\s*(\\d+) tokens"        # regex with one capture group
      tokens_out: "Output:\\s*(\\d+) tokens"
      cost: "Total:\\s*\\$([\\d.]+)"
      llm_calls: "(\\d+) API calls"

No patterns? agentarena still measures wall time and pass/fail — works for any tool.

Examples

See examples/ for ready-to-use configs:

• python-pytest.yaml — Python with pytest, mypy, bandit
• node-typescript.yaml — TypeScript with tsc, jest, ESLint
• react-nextjs.yaml — Next.js with vitest
• go.yaml — Go with go test, race detector

Contributing

See CONTRIBUTING.md for development setup and guidelines.

License

MIT