agent-bench

[]

Benchmark AI coding agents against each other on the same task.

Run the same coding task across Claude Code, Codex CLI, Gemini CLI, Aider, OpenClaw, and more — then compare cost, speed, token usage, code quality, and test pass rate.

Why?

No existing tool lets regular devs compare AI coding agents head-to-head on their codebase. Terminal-Bench is academic. This is practical.

Install

pip install cli-agent-bench

Quick Start

# Create config
agent-bench init

# Edit .agent-bench.yaml with your task and agents

# Run all configured agents
agent-bench run

# Run specific agents
agent-bench run --agent claude-code,codex-cli

# Override the task
agent-bench run --task "Add error handling to all API calls"

# View results
agent-bench results

# JSON output
agent-bench results --json

# View history
agent-bench history

# Check which agents are installed
agent-bench agents

Example Output

╭─────────────────── Agent Benchmark Results ────────────────────╮
│ Task: "Add pagination to users endpoint"                        │
│ Run: 2026-04-07 00:15                                          │
├──────────┬──────────┬───────┬──────────┬────────┬──────────────┤
│ Agent    │ Time     │ Cost  │ Tokens   │ Tests  │ Quality      │
├──────────┼──────────┼──────────┼────────┼────────┼──────────────┤
│ Claude   │ 2m 14s   │ $0.42 │ 18.2K    │ 8/8 ✅ │ A (92/100)   │
│ Codex    │ 1m 47s   │ $0.31 │ 14.1K    │ 8/8 ✅ │ A- (88/100)  │
│ Gemini   │ 3m 02s   │ $0.18 │ 22.3K    │ 7/8 ⚠️ │ B+ (81/100)  │
│ Aider    │ 4m 15s   │ $0.55 │ 31.0K    │ 6/8 ❌ │ B (75/100)   │
╰──────────┴──────────┴───────┴──────────┴────────┴──────────────╯

Winner: Claude Code (A — best quality)
Fastest: Codex CLI (1m 47s)
Cheapest: Gemini CLI ($0.18)

Configuration

.agent-bench.yaml:

agents:
  claude-code:
    command: claude
    args: ["--dangerously-skip-permissions"]
  codex-cli:
    command: codex
    args: ["--full-auto"]
  aider:
    command: aider
    args: ["--yes-always"]

default-task: "Refactor this file to use type hints throughout"

scoring:
  run-tests: true
  lint: true
  timeout: 300

Quality Score

Component	Weight
Test pass rate	40%
Lint clean	20%
Code diff sensibility	15%
Task completion	15%
Speed bonus	10%

How It Works

1. Copies your project to an isolated temp directory per agent
2. Runs each agent as a subprocess with the task prompt
3. Captures stdout/stderr, exit code, duration
4. Parses token usage from output
5. Runs tests and linter if configured
6. Calculates cost from token usage + model pricing
7. Scores quality across multiple dimensions
8. Stores results in SQLite for history
9. Cleans up temp directories

Supported Agents

Any CLI-based coding agent: Claude Code, Codex CLI, Gemini CLI, Aider, OpenClaw, Hermes, OpenCode, and more. If it runs in a terminal, it works.

License

MIT © Hiren Thakore