1 person Hermes Agent 跑整個行銷團隊：4 層拆解
脈報 · 2026-05-18 · 8 min
Show full episode description
AI marketer Shann 把整套行銷流程搬到 Hermes Agent 上跑。我拆解 4 層架構、4 個 Level 與 prototype 到 production 四步法，順便講即使不用 Hermes 也值得學什麼。 ⭐ 文章深度讀：完整 4 層架構與作者沒回的成本反例都寫在文章裡，看完想留言聊聊也歡迎 → https://heymaibao.com/hermes-agent-marketing-team-4-layer-setup/ ⚡ 章節重點 開場
✨ Episode Outline — click any point to jump to it in the episode
Problem solved
How a solo AI marketer scales into an automated multi-agent team using Hermes Agent's four-layer architecture.
Benefits
- Out-of-the-box preloaded skills save setup time
- Agent accumulates experience and gets smarter over time
- One brain reused across six distinct personalities
- Governance separated from execution for safe scaling
- Right model assigned to the right task type
Use cases
- Transform a single-handed marketing workflow into a fully automated invisible fleet
- One user spent $300 in API fees in just three days testing the framework
- Hand creative copywriting tasks to Cloud Over 4.7, structured coding to Codex or GPT 5.5
- Scale from Level 1 solo assistant to Level 4 fully automated cloud fleet via VPS scheduling
KPIs / results
- $300 in API fees in three days
- Four-layer architecture, four roles, four Levels, four deployment steps
📑 Chapters — tap a time to jump there
00:00
01:20
第 1 層：單一 Agent 的組成
- Single agent = three blocks: brain, personality, skill set
- Brain stores stable facts across conversations
01:43
第 2 層：4 個角色的心智模型
- Four-role mental model separates governance from execution
- Rebuild body from brain, not brain from body
02:57
第 3 層：4 個 Level 的演進路徑
- Four Levels: solo to automated cloud fleet
- Level 3 adds front-door dispatcher; isolation is king
04:07
第 4 層：prototype 到 production 四步法
- Four steps: prototype, auto-generate skills, isolate workspace, VPS schedule
- Models split: Cloud Over 4.7 vs Codex/GPT 5.5
05:38
現實檢驗：成本與時間表的縫隙
- Reality check: $300 API fees in three days
- Steep Docker/SSH learning curve, unverified data claims
06:33
收尾與我的判斷
- Powerful framework but basic fine-tuning work is indispensable
🌐 This transcript was automatically translated to English from the original.
Well, if you happen to be a content creator, or someone who engages in AI marketing every day, you will definitely love this episode. Just like when we were drinking coffee in private, I can’t wait to share with you a super powerful framework that I just discovered today. We are going to dismantle how Shan, an AI marketing expert, relied on a thing called Hermes Agent to transform his original single-handed workflow into a fully automated invisible fleet. Are you ready? Let’s get right to the point. Let's first listen to Shan's own definition. He said that this is an autonomous agent that will become stronger with longer running time. Wow, this sentence directly raises our expectations today, right? Think about it, this is not the kind of stupid robot that you have to submit again every time you have a conversation. It is more like a digital employee who can really accumulate experience, and the more you use it, the smarter it becomes. Okay, let me give you my judgment first. Why should you choose Hermes? In fact, the key is that its built-in default values can produce huge negative effects. Compared with the Linux Pi approach that requires you to issue explicit instructions for everything, Hermes takes an out-of-the-box Rails Pi route. It directly preloads one, two or three skills for you. In other words, before you even start writing the first line of settings, it has already saved you a lot of time. Therefore, our analysis today will take you through a rigorous four-layer architecture. The first layer is a single Agent. The second layer is the mental model of the four characters, then the third layer is the path of his glasses, and the fourth layer is the deployment steps. Of course, in the end we will do a super pragmatic reality check and calculate how much it will cost. Okay, let’s look at the first layer right away, the composition of a single Agent. Let’s dissect what a Hermes worker looks like. In fact, every Hermes has three core blocks. The first is the brain, which will store stable facts in the memory. Let him remember your background across conversations. Next is personality, which determines the tone and atmosphere of his speech. Finally, of course, there is the skill set, which includes the one, two, and three preset skills just mentioned, as well as the abilities he has learned by himself. Here is the coolest part, you can use the same brain to match six completely different personalities. In an instant, he can switch from a dynamic business person to a super rigorous researcher. It's amazing. However, Shan gave a super counter-intuitive suggestion here. I think this must be an important point. He said, absolutely, never write skills by yourself on the first day. Why? Because you should first let the Agent observe your real work process. If you intervene too early, you will actually just write your current biases and those inherently inefficient bad habits directly into the system. Then we enter the second layer, the mental model of the four roles. This can be said to be the most valuable architectural secret in the entire framework. He perfectly demonstrates how governance and execution must be completely separated. You may ask, What is Agent Control? To put it simply, it is a folder used purely for management. It determines which Agents can exist, what rules they have to abide by, and their respective permissions. This is strictly separated from the real-time running entities. Only by separating the management of the brain from the people who perform the work on the front line can you prevent your system from becoming an unmanageable disaster when it expands. There is a sentence that perfectly sums up this philosophy. You can rebuild the body from the brain, but you cannot rebuild the brain from the body. This means that this control mode is the brain of your system. The running Agents outside are just bodies that can be replaced at any time. Just like the Agent on the front line today has completely collapsed. As long as the logic in your control is still alive, you can restart a new one at any time without panic. Next is the third level, the glasses path. Let's see how you can expand from a small assistant to an entire powerful virtual department. There are four levels here, from Level 1 where you work alone all the way to Level 4 of a fully automated cloud fleet. But I hope you pay special attention to Level 3. Because at Level 3, you will join a role called the front door dispatcher. This is a watershed, and it is the decisive moment when your originally scattered Agents truly transform into a cohesive team. Therefore, the core principle is that isolation is king. You must avoid the so-called super Agent trap. I know, it seems very convenient to give all memories and permissions to the same entity, but if you do this, he will soon become confused and confused about his identity. And by that time, If you want to take back a specific permission, sorry, the system simply can't do it. As for model selection, Shan's division of labor strategy is very smart. He will hand over heavy tasks that require creativity, copywriting, or taste judgment to Cloud Over 4.7. If it is the kind of structured task that requires step-by-step programming and high predictability, then leave it to Codex or GPT 5.5. This is called letting the right model do the right thing. OK, let's come to the fourth layer. The number of steps from the circle to the official launch. This four-step method is super practical. I dare say, you can just steal it and apply it to any AI framework you have. First, you need to build a circle on the main Hermes to try and error. Second, use the real work content two or three times to let it automatically generate skills. Then, the third step, isolate it to a dedicated workspace for fine-tuning. Then, throw it to the VPS cloud to set the schedule. It is really just like what Jen said. It is impossible to print an Agent from scratch that can be directly launched. You must slowly develop it through these four steps. However, before you rush to cut off all existing workflows and re-train, wait a minute, we need to do a pragmatic reality check without filters. 300 US dollars. Everyone, this is a very shocking number. One user jumped out and complained that he spent 300 US dollars in API fees in just three days just to test this framework. The point is, Shan has yet to give a positive response to the question of this ultra-high cost. To be honest, in addition to the cost being a million dollars, there are other blind spots. For example, the ultra-strong data he claimed have not been independently verified, and the schedule that can be built in a few weeks is a bit too radical. Not to mention that if you want to understand Docker and SSH, the learning curve is quite steep. Of course, this is not to say that this framework is not easy to use, but when you listen to others enthusiastically recommending tools, these are the key realistic backgrounds you must have. I won’t blindly follow the trend. So in the end I want to leave a question for everyone to think about. Before you start fantasizing about having a Level 4 fully-automatic armada, are you really willing to go through the tedious and boring fine-tuning process of Level 1 every day? This set of tools is indeed very powerful, but the basic work is definitely indispensable. If you want to be lazy and skip the basics, you will definitely hit a wall during deployment. Okay, that’s it for today’s analysis. I hope it will be helpful to you. See you next time. Please ask.