agentleFS
Sign inStart free
28672032962867203296/hermes-agent/skills/agent-evaluation-framework/SKILL.md

agent-evaluation-framework

Skill0 starsChanged 3 months agoMITView on GitHub

Agent 评估框架 v2：自动回归测试+A/B对比+SWE-Bench实战。借鉴 OpenAI 噪声分离 + Databricks 百万行基准 + Google A2A eval

1 other project uses a version of this.

Adopted by

Most recently taken up by

The file

Create a free account to mount
Files
SKILL.md

Agent 评估框架 v2.0

v1.0：手动评估。v2.0：自动回归 + A/B 测试。

一、自动回归测试（v2.0 新增）

每次改配置/换模型/更新技能后，自动重跑 10 个代表性任务：

回归测试集

D:/Hermes/eval/regression/
├── L1_open_file.json       "读config.yaml，改timeout值"
├── L1_search_code.json     "搜所有含'MCP'的py文件"
├── L2_fix_bug.json         "TTS播报失败→诊断→修复→验证"
├── L2_add_feature.json     "给skill加一个triggers条目"
├── L2_research.json        "查GitHub trending AI项目，输出报告"
├── L3_refactor.json        "重构一个模块，跑测试，无回归"
├── L3_full_feature.json    "从需求到PR的完整功能"
├── L4_explore.json         "调研一个新MCP服务，写接入方案"
├── edge_empty_input.json   "边界：空输入处理"
├── edge_large_codebase.json "边界：500文件代码库搜索"

自动运行

python D:/Hermes/eval/run_regression.py --all
# 输出：8/10 PASS, 1 FAIL (L2_fix_bug: 步骤3超时), 1 SKIP

回归告警

连续 3 次失败 → 标记为 regression → 阻断后续升级直到修复

二、A/B 对比测试（v2.0 新增）

对比维度

对比项A组B组评估指标
模型DeepSeek V4Opus 5成功率+token成本
Skill有 semantic-cache无重复查询token消耗
策略SigMap定位grep定位定位准确率+token
Prompt原则式(5条)规则式(16条)任务完成率

A/B 测试模板

# D:/Hermes/eval/ab_test.py
tasks = load_tasks("regression/*.json")
results_a = run_tasks(tasks, config="A")
results_b = run_tasks(tasks, config="B")
compare(results_a, results_b)  # 输出胜者+置信度

三、成本归因（v2.0 新增）

不只是"这次花了多少token"，而是"这笔token花在哪个阶段"：

任务: 查GitHub trending
├─ 搜索阶段: 1,200 token (15%)
├─ 读取阶段: 3,800 token (47%)
├─ 分析阶段: 2,100 token (26%)
└─ 输出阶段: 1,000 token (12%)
总: 8,100 token | $0.04

（v1.0 保留的五维评估卡、四类测试任务、评估原则不变）

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

Check your setup, free