with one click
openbench
Agent A/B 测试平台 — 对比不同 agent 配置(模型、提示词、工具、轮次),统计显著性评估最优方案。
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Menu
Agent A/B 测试平台 — 对比不同 agent 配置(模型、提示词、工具、轮次),统计显著性评估最优方案。
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
Control message formatting and writing style for Lark/Feishu conversations.
Create, query, edit, and manage Feishu Bitable (multidimensional spreadsheets) via lark-cli base — apps, tables, fields, records, views, dashboards, forms, workflows.
Manage Feishu calendar events via lark-cli — create/update/delete events, manage attendees, query free/busy and suggest available time slots, find and book meeting rooms.
List Feishu group chat members and resolve chat metadata via lark-cli.
Create a Feishu cloud document from Lark-flavored Markdown via lark-cli, with optional folder or wiki placement.
Comment management and media (image/file) operations on Feishu cloud documents via lark-cli.
| name | openbench |
| description | Agent A/B 测试平台 — 对比不同 agent 配置(模型、提示词、工具、轮次),统计显著性评估最优方案。 |
| triggers | {"intent_patterns":["benchmark|A/B|ab test|对比测试|openbench|tournament|实验"],"context_signals":{"keywords":["benchmark","A/B","openbench","对比","实验","tournament"]},"confidence_threshold":0.7} |
| priority | 6 |
| requires_tools | ["bash"] |
| max_tokens | 200 |
| cooldown | 60 |
| enabled | false |
| disabled_reason | Depends on the external openbench CLI/repo and is not self-contained in this workspace. |
Agent A/B 测试平台。自动化 plan → run → evaluate → repeat,找到最优 agent 配置。
openbench 支持单变量 A/B 实验、多 agent 循环赛、以及自然语言驱动的自动研究循环(LLM 生成假设 → 执行 → 评估 → 迭代)。
| 场景 | 示例 |
|---|---|
| 模型对比 | haiku vs sonnet 在编码任务上的正确率和成本 |
| 提示词优化 | 简洁 vs 结构化 system prompt 的效果差异 |
| 工具配置 | 有无 Bash 工具对数学题正确率的影响 |
| 轮次调优 | max_turns=5 vs 20 的完成率和 token 消耗 |
| 自动研究 | 用自然语言描述优化目标,LLM 自动生成并执行实验 |
openbench run experiments/my_test.py
openbench tournament experiments/my_tournament.py
openbench research "找到在编码任务上性价比最高的模型+提示词组合" --max-iter 5 --max-cost 10
openbench compare experiment_name
openbench list # 列出所有实验
openbench show experiment_name # 查看详细结果
openbench runs experiment_name # 列出所有运行
openbench lineage experiment_name # 版本演进追踪
openbench tui # 交互式浏览器
from openbench.types import AgentConfig, DiffSpec, Experiment, TaskItem
experiment = Experiment(
name="prompt_style",
description="简洁 vs 详细 system prompt",
diff=DiffSpec(field="system_prompt", description="prompt verbosity"),
agent_a=AgentConfig(
name="minimal",
model="claude-haiku-4-5",
system_prompt="Be concise.",
allowed_tools=["Bash"],
max_turns=10,
),
agent_b=AgentConfig(
name="detailed",
model="claude-haiku-4-5",
system_prompt="You are a senior engineer. Think step by step...",
allowed_tools=["Bash"],
max_turns=10,
),
tasks=[
TaskItem(prompt="Write fizzbuzz", expected="1,2,Fizz,...", check_fn='"Fizz" in output'),
],
num_samples=3, # pass@k
)
| 参数 | 说明 |
|---|---|
--dry-run | 预览不执行 |
--samples N | 每 (agent, task) 的试验次数 |
--max-iter N | 自动研究最大迭代数 |
--max-cost $X | 预算上限(USD) |
--model MODEL | Claude 模型 |
--target | 优化目标:quality / cost / latency |
--yes | 跳过确认 |