编码代理(Claude Code、Aider、Codex等)在自定义任务上的直接比较,包含通过率、成本、时间和一致性指标
一个轻量级 CLI 工具,用于在可复现的任务上对编码代理进行头对头比较。每个“哪个编码代理最好?”的比较都基于感觉——本工具将其系统化。
# pinned to v0.1.0 — latest stable commit
pip install git+https://github.com/joaquinhuigomez/agent-eval.git@6d062a2f5cda6ea443bf5d458d361892c04e749b
以声明方式定义任务。每个任务指定要做什么、要修改哪些文件以及如何判断成功:
name: add-retry-logic
description: Add exponential backoff retry to the HTTP client
repo: ./my-project
files:
- src/http_client.py
prompt: |
Add retry logic with exponential backoff to all HTTP requests.
Max 3 retries. Initial delay 1s, max delay 30s.
judge:
- type: pytest
command: pytest tests/test_http_client.py -v
- type: grep
pattern: "exponential_backoff|retry"
files: src/http_client.py
commit: "abc1234" # pin to specific commit for reproducibility
每个代理运行都获得自己的 git 工作树——无需 Docker。这提供了可复现的隔离,使得代理之间不会相互干扰或损坏基础仓库。
| 指标 | 衡量内容 | |--------|-----------------| | 通过率 | 代理生成的代码是否通过了判断? | | 成本 | 每个任务的 API 花费(如果可用) | | 时间 | 完成所需的挂钟秒数 | | 一致性 | 跨重复运行的通过率(例如,3/3 = 100%) |
创建一个 tasks/ 目录,其中包含 YAML 文件,每个任务一个文件:
mkdir tasks
# Write task definitions (see template above)
针对你的任务执行代理:
agent-eval run --task tasks/add-retry-logic.yaml --agent claude-code --agent aider --runs 3
每次运行:
生成比较报告:
agent-eval report --format table
Task: add-retry-logic (3 runs each)
┌──────────────┬───────────┬────────┬────────┬─────────────┐
│ Agent │ Pass Rate │ Cost │ Time │ Consistency │
├──────────────┼───────────┼────────┼────────┼─────────────┤
│ claude-code │ 3/3 │ $0.12 │ 45s │ 100% │
│ aider │ 2/3 │ $0.08 │ 38s │ 67% │
└──────────────┴───────────┴────────┴────────┴─────────────┘
judge:
- type: pytest
command: pytest tests/ -v
- type: command
command: npm run build
judge:
- type: grep
pattern: "class.*Retry"
files: src/**/*.py
judge:
- type: llm
prompt: |
Does this implementation correctly handle exponential backoff?
Check for: max retries, increasing delays, jitter.
Copy a source-pinned command for your client. You run it yourself.
Destination: .claude/skills/agent-eval · pinned to the source commit
git clone https://github.com/affaan-m/ECC.git
cd ECC
git checkout d8409a4b0813771235555e32e3d8046a73988bfa
mkdir -p ".claude/skills/agent-eval"
cp -r "docs/zh-CN/skills/agent-eval" ".claude/skills/agent-eval"Review the source before running. This copies files into your project; it is not a one-click install and does not verify runtime safety.
Scanner static-checks@0.1.0 · commit d8409a4b0813. Static checks cannot prove runtime safety – review the source and the exact diff before installing. How checks work.
Instructs shell/process/package operations that run commands on the host.
Evidence: pip install· fingerprint 7944ec554efca445
Uses encoding/eval patterns that can hide executable content.
Evidence: [redacted]· fingerprint 7f712a33096ba821