Agent capability comparison methodology

Skill ChrisLamDev/hermes-core-skills/skills/agent-capability-comparison-methodology

25 executable AI agent skills for debugging, planning, token efficiency, and securityFrom the repository description

Install

npx -y skills add ChrisLamDev/hermes-core-skills --skill agent-capability-comparison-methodology

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

2 things to look at

10 stars10 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
runs commandsInstructs the agent to run 2 commands, including `ls -d */` and 1 more.

SKILL.md

4.4 KB, ~1.5k tokens by cl100k_base, as published. Nobody here has run it

AI Agent 能力比較方法論

適用場景

用戶想比較兩個 AI agent/tool 嘅能力差異，例如「Hermes Agent vs OpenAI Codex」、「CUA vs Codex Desktop Control」。

核心原則

唔靠「聽講」，用證據說話。 三層驗證：

官方文件 — GitHub README、docs 目錄、官方網站
原始碼證據 — 直接睇 source code 確認功能存在與否
實測驗證 — 真係裝黎用、試下得唔得

比較流程

第一層：快速掃描（30 分鐘）

睇 GitHub README 了解定位
睇 docs/ 目錄了解功能列表
記錄 GitHub stats（stars、forks、commits、issues、branches、tags）

第二層：深入驗證（1-2 小時）

逐個功能查原始碼：

功能類別	要 check 嘅嘢
記憶系統	source code 有冇 memory/memories 目錄？點樣跨 session？
Cron/排程	有冇 scheduler/cron crate/module？
多平台支援	有冇 telegram/discord/slack module？
模型自由度	model provider 係硬編碼定可配置？
Computer Use	有冇 screenshot/click/type_text 工具？
MCP 整合	有冇 mcp server/client 實作？
Skills	係 static 定 dynamic learning？
Desktop App	係咪真係有 Electron/Tauri app？

檢查方法：

ls -d */ 睇目錄結構
grep -r "keyword" --include="*.py" --include="*.rs" --include="*.ts" 搜尋關鍵字
睇 GitHub Issues 用 label 過濾（如 computer-use、memory、cron）
睇 Discussions 嘅 feature requests

第三層：實測驗證（按需要）

真係裝黎用
試核心功能
紀錄成功/失敗嘅 case

關鍵：Always distinguish CLI vs Desktop App features

Codex Desktop has Computer Use + Automations; Codex CLI does NOT
Hermes CLI has cron/scheduling; no official Desktop App yet
Check which variant each feature belongs to before comparing

關鍵：Memory systems may be read-only

Codex has a two-phase memory pipeline (rollout extraction + global consolidation)
But the agent is explicitly instructed: "Never update memories. You can only read them."
Hermes has bidirectional memory — agent can both read AND write
This is a meaningful functional difference

關鍵：「Desktop App」嘅定義要查原始碼

Codex「Desktop」可能只係 npm CLI 包裝，唔係 Electron/Tauri GUI
Hermes 嘅 Electron app 可能已存在但未公開（apps/desktop/ 目錄）
要 check：有冇 electron/、tauri/、gui/ 目錄，或者睇 package.json/CMakeLists.txt

陷阱 5：Skills 系統是否自動學習？

Some agents have static skills (user writes SKILLS.md, never changes)
Some agents auto-create skills from experience and improve them during use
Check the source code for auto-learning logic vs static definition

陷阱 6：Model Provider 支援要查原始碼定義

README 可能冇列出所有 provider
Check model-provider crate/module 嘅 enum/struct 定義
Codex supports: openai, amazon-bedrock, ollama, lmstudio, and custom OpenAI-compatible
Hermes supports: 200+ via OpenRouter + direct OpenAI, Anthropic, DeepSeek, Ollama, etc.

常見陷阱（實測教訓）

陷阱 1：「Desktop App」嘅定義

Codex「Desktop」可能只係 npm CLI 包裝，唔係 Electron/Tauri GUI
要 check：有冇 electron/、tauri/、gui/ 目錄，或者睇 package.json/CMakeLists.txt

陷阱 2：「有 feature request = 有功能」

GitHub issue 話「add X feature」唔代表已經實作
要 check：source code 有冇相關實作，issues 係咪 open/closed

陷阱 3：Tag 數量 = Release 頻率？

1,045 tags 可能包含大量 pre-release/CI tags
要 check：實際 release notes、正式版 vs pre-release 比例

陷阱 4：官方網站需要 login

developers.openai.com/codex 回傳 403
呢啲情況要 rely on source code

報告格式

用表格對比，每個功能一行：

功能	Agent A	Agent B	贏家
功能名	✅ 有 / ❌ 冇	✅ 有 / ❌ 冇	A/B/平手

最後加一段總結，講清楚：

邊個適合咩場景
最重要嘅差距係咩
有冇誤解需要澄清

What ships with it

Read from the repository

Just SKILL.md. No reference files, no scripts.