AI agent 不是銀彈：拆穿 YouTube 最愛騙你的 6 個謊言
脈報 · 2026-06-02 · 8 min
Show full episode description
YouTuber Craig Hewitt 拆穿 always-on AI agent (Hermes、OpenClaw) 的 6 個迷思，從要花多少錢買電腦到「整個事業靠 agent 跑」的幻覺，幫你判斷它真正該用在哪。 ⭐ 文章深度讀：影片拆了 6 個迷思，但「哪些事該丟給 always-on agent、哪些該自己坐在鍵盤前做」這條判準，我在文章裡用一句話加實例講得更細 → https://heymaibao.com/ai-agent-six-lies/ ⚡ 章節重點 開場
📑 Chapters — tap a time to jump there
00:00
先記住這條判準
01:01
硬體沒那麼貴
01:26
便宜訂閱不夠用
02:49
事業全靠 AI 是幻覺
03:53
養一個 agent 的務實法則
04:42
它是隊友不是替代品
06:03
要不要用的三步測試
07:11
收尾
🌐 This transcript was automatically translated to English from the original.
Hello everyone, do you often see people showing off on the Internet recently, saying that I use a whole team of AI agents to fully automate the entire career? Do you feel like you are lagging behind, or even a little anxious about information? Really, don’t be nervous. For today’s analysis, we will use the practical experience of expert Craig Hewitt to pull the Always on AI agent, that is, the kind of AI agent that is always online, from the over-touted magic tank. We will use the most pragmatic perspective to help you crack it one at a time. 6 of the most common AI myths Are you ready? Let’s get right to the point. Have you ever felt like this? You just feel like you have to spend $5,000 to get started with AI automation. You need to buy a military-grade Mac Studio or a top-of-the-line computer to run it. Today we’re just chatting with friends. Let’s start with a cruel reality tea box. Do you really need to spend this big money to get started? No kidding. Let’s break it down one by one. 6 Before this myth, let’s talk about the conclusion. This is the only gold standard today that you need to remember. Does this task really require your physical presence? To put it bluntly, as long as you draw this line clearly, you can truly amplify your productivity instead of being blindly kidnapped by new tools. Okay, let’s take a look at the first part. Reality tea box, myths about hardware and cost. Many people feel that they have to run to a framework like Hermes that operates silently in the background. The threshold must be scary high. But when we look at it, the most pragmatic hardware options are actually these three. First, the advantage of cloud VPS is that it is always online and only costs 5 to 10 US dollars per month. However, because it is not a general stationed IP, if you want to use it to crawl Reddit, it is very easy to be blocked, and there are security risks. Second, buy a Mac Mini and run it on the ground. The IP is clean, and you can use Tail Scale to strengthen security. The disadvantage is that you have to pay a sum of money upfront. However, please pay special attention to the third option, which is also the one I definitely recommend to novices. It’s the ordinary computer you’re using now. It’s free. It’s the easiest. Just start here. Not to mention Toy One, even if you really want to upgrade the hardware and get a machine that operates independently and doesn’t interfere with your work at all. The actual number is 600. Buying a Mac Mini is almost 600 US dollars. This is really 108,000 US dollars different from the myth of 5,000 US dollars spread on the Internet. So, the hardware cost It definitely shouldn’t be an excuse for you not to start. After you get the hardware, the next big myth is the cost of the model. Many people think that they can just bind the cheapest Chat GPT plan and run without thinking. Practical experience will slap you in the face. You will definitely be restricted from the plan that costs 20 US dollars per month. If you are serious about running Always on Agents, upgrade to the 100 US dollars plan. It is really powerful. So how do real experts mix and match tools? You see, Crack's model pants are like a weapons shop. Leave real-time network research to Perplexity or Grog. Leave in-depth planning and architecture to Opus 4.8. Use GPT-6 or Codex for daily work. Leave lightweight background tasks to Hiku or even super cheap ones. Just leave it to Tremany 3. Others like Kimi, Minimax or Deepsea. They are also super easy-to-use weapons. It will even modify the Soul.MD file in the system to make the tone of the model's speech sound more relaxed and natural. So you see, there is no universal model. Picking the right tool for a specific task is the way for an expert to play. Next, we are going to talk about the second part, exposing the biggest illusion that the entire business depends on AI. We often see people bragging on the Internet that all marketing, business, and operation teams have been replaced by AI overnight. In response to this pink bubble, SARS expert Jason Lamkin and Craig Bellemus proposed a super-accurate metaphor called the 80-80 rule. What does it mean? It only takes 80% of your efforts to make an agent into a super cool demo video. It's easy to give people the illusion of "Wow! I'm almost done!" But the cruelest reality is that if you want to make it stable and customized enough to be put into daily operations, you need to spend another 80% of your efforts. To be honest, just maintaining it every day without making mistakes is a full-scale job in itself. So instead of dreaming about raising an entire AI army there, Why not obey these pragmatic rules? First, you are not forming a rock band. Please don't think about raising a whole team from the beginning. Just focus on raising an agent within six months. Second, don't anthropomorphize it. Giving it a super-killing name will not make it stronger. Third, this is super important. Be sure to set up safety guardrails and turn off the automatic skill creation function to prevent your system from becoming extremely bloated. Finally, never automatically install skill files from unknown sources and those seemingly harmless markdown files. There may be huge risks of supply chain attacks hidden in it. Speaking of system bloat, many users have recently switched jobs from OpenCloud to Hermes. The key lies in the memory view. Hermes' system will automatically prune and compress knowledge, but Open Cloud can easily lose the context during updates. This once again proves one thing. More skills and techniques are definitely not better. Having a few skills that you can fully trust and that are highly consistent with your work context is far more important than hoarding a lot of slippery and unrealistic functions. Okay, after dispelling the myths, let’s talk about the most pragmatic division of labor in the third part. Is AI your opponent or a teammate? In fact, if you look at the always on agent like Hermes, what he is best at is continuous asynchronous tasks, just like building blocks stacked and falling infinitely in the background. However, Cloud Code is written as a collaborative tool. It requires you to be present and iterate over and over again to shape your ideas bit by bit. To be honest, for 95% of people, this is a tool that requires your deep participation. At present, it is still your most powerful force to significantly improve productivity. They are teammates, not competitors. So what is Always on agent best suited for? I will give you the most perfect example. Use Amplify to crawl Reddit eight times a day, and then sort out the potential customer posts for me that I can reply to. You see, this is an extremely specific and highly revisited task, and it is super suitable for tasks that are performed silently in the background. This is the comfort zone where Agent should really stay. Finally, the fourth part is the ultimate practical test to decide whether you need Agent or not. We have directly condensed a lot of concepts mentioned above into these three no-brainer steps. You can check your workflow immediately after listening. The first step is to find an agent. The second step is to ask yourself: Do I need to be there to shape and lead this matter? The third step. If the answer is yes, then no problem. Open Cloud Code and do it yourself. If the answer is no, it just needs to be executed silently in the background. Then boldly hand it over to Hermes or OpenCloud. It’s that simple and crude. Okay, now open your to-do list for today. Ask yourself carefully what things here really require you to sit in front of the keyboard to complete. Really don’t be anxious about the exaggerated AI hype outside. Use the right framework to allocate the right tools. Today’s explanation is shared here. I hope this can help you completely free up your brain and time to do things that can really generate value. See you next time. Okay.