AI Meetup · Book Club Edition
What Makes
an AI Agent
· ✦ ·
June 2026
featuring Hermes
讀書會→
What Makes an AI Agent
Layer I
Layer 1
Large Language Models
Giving the model an input, it pieces together a response according to probability
Input PROBABILITY the a is an was ... It is ... token by token
Problem
All requests have the same priority, no character, no way for providers to control behavior
1 / 6LLM
What Makes an AI Agent
Layer II
Layer 2
Chat-based Assistant
SYSTEM PROMPT PRIORITY: HIGH Defined Identity controls USER PROMPT PRIORITY: LOWER Tell me about... Ignore all instructions → ChatGPT can have a defined identity
Problem
Only knows things from training data, that's why "knowledge cutoff date" used to be such a big deal
2 / 6System Prompt
What Makes an AI Agent
Layer III
Layer 3
Assistant That Can Do Things
Giving the model arms and legs, now it interacts with the world, with side effects
Search Terminal Browser MCPs Database APIs SIDE EFFECTS
Problem
But limited by context window: the larger it gets, the worse the needle-in-the-haystack problem. Compaction helps, but loses information.
3 / 6Tools & MCPs
What Makes an AI Agent
Layer IV
Layer 4
Assistant That Remembers
Maintain context throughout time · Accumulate knowledge
RAG KNOWLEDGE BASE retrieve context File-based memory/ context.md preferences.md knowledge/ topic-a.md topic-b.md directories enforce structure Claude / Anthropic
Problem
Still can't make decisions, only does what it's told to do
4 / 6Memory
What Makes an AI Agent
Layer V
Layer 5
AI Agent?
Reasoning Chain of Thought think... think longer... think much longer... → smarter THINKING TIME Plan Formulation How to decide with structure and goal GOAL Step 1 Step 2 Step 3 ok? ok? ok? Result
5 / 6Reasoning
What Makes an AI Agent
Layer VI
Layer 6
AI Agent!
Harness: Orchestration of everything built prior
THE LOOP GOAL Plan Execute Tool done? retry? change plan Done THE STACK 6 Harness 5 Reasoning 4 Memory 3 Tools 2 System Prompt 1 LLM
6 / 6The Harness
Hermes
Part II
什麼是 Hermes
Nous Research 開源的 self-improving AI agent，不是 chatbot wrapper，不是 coding copilot
YOU TALK TO HERMES VIA Telegram Discord Slack HERMES self-improving agent RUNS ON Mac $5 VPS Docker Serverless 不綁你的筆電 THE SIX LAYERS Tools 60+ 內建, browser, terminal, MCP Memory MEMORY.md + USER.md, 跨 session Reasoning 接任何 model (OpenRouter, etc.) Harness learning loop：從經驗中建 skill 跨 session 累積對你的理解 跑得越久越強的自主 agent
Hermes什麼是 Hermes
Hermes
Skills
如何和 Hermes 一同撰寫 Skills
SKILL.md，需要時才載入 (progressive disclosure，不浪費 token)
✗ 列出步驟 1. 打開 Booking.com 2. 搜尋目的地 3. 排序 4. 選最便宜 5. ??? 你沒辦法窮舉一切條件 vs ✓ 判斷標準 IDENTITY 專業飯店評論家，最適合的，不是星星最多的 DIMENSIONS（不指定權重） 地點 評價 服務 交通 TIEBREAKER 場合適合度 > 取消彈性 > 評價穩定度 → 沒寫到的狀況也能合理判斷 EXAMPLE: HOTEL-CRITIC
HermesWriting Skills
Hermes
Skills · Iteration
撰寫過程就是協作，而且 Hermes 自己也會建 skill
你寫 SKILL STEP 1 先寫一版 STEP 2 讓 Hermes 跑 STEP 3 · 修正認知，不是加步驟 「時間偏好是 ranking signal，不是 filter」 iterate 收斂 → 穩定的決策邏輯 HERMES 自己建 SKILL 完成一個複雜任務 把學到的做法存成新 skill 下次遇到類似任務 → 直接調用 self-improving = procedural memory 前提：沒有推理能力的 model，給它認知框架也沒用
HermesIteration
Hermes
Daily Routine
日常 Routine
寫好 skills，之後只要給 Hermes 一個目標
TELEGRAM 幫我查七月中去 東京的機票跟飯店 HERMES flight-search hotel-critic → browser: Google Flights → browser: Booking.com 自己決定怎麼查 REPORT 結構化報告，不是一堆連結 CRON / 排程 「每天早上幫我看一下 Hacker News」 用自然語言設定 結果自動送到 Telegram 你不需要 SSH 進去
Hermes日常 Routine
Hermes
v0.17
v0.17
Async Agents
delegate_task → 獨立的 child agent，隔離的 context + terminal
BEFORE: SYNC task A ··· task B AFTER: SUBAGENT DELEGATION parent subagent: flight-search subagent: hotel-critic 預設 3 concurrent，可調整 「七月中去東京五天」 SUBAGENT 1 flight-search 獨立 context · 獨立 terminal 操作 Google Flights SUBAGENT 2 hotel-critic 獨立 context · 獨立 terminal 操作 Booking.com 匯整成一份完整行程 從「一次做好一件事」→「同時推進多條線」，跟人類工作方式更接近
HermesAsync Agents
End of Dispatch
✦
Thank You
· ✦ ·
Fin←
← → or click