I have installed Clawdbot (now OpenClaw, also known as Moltbot) on my Mac Mini, connected it to Telegram, extended its abilities such as making it able to read and write emails, to alert extreme weather changes, summarize my kids' school activities, read a list of newsletters and write weekly blogs to track AI industry news, connect it to my home camera and let it parse detected events and trigger alerts. The most common usable is to let it config setup itself for all these project. Yes, with the bash tool and the control of the machine, it can not only help me code up all the functions above, but also can configure itself for everything. All I need to do it tell it what I want to do. Only describe the desired outcome and it will try to figure it out how to implement. I would like share a few thoughts about this product.

It is very different to chat bots such as Claude and ChatGPT. The pure chatting has its role nowadays. But not able It is also different from coding agents such as Claude Code and ChatGPT Codex. Claude Code set customer mindeset as a coding agent, although I think it can do more than just coding. But once the mindset is formed, it is hard to change. Clawdbot set intial customer mindeset as a all-round general personal assistant. This mindset is very important. I have point it out that there is no product with this mindeset from leading AI companies, yet. OpenAI Atlas is an interesing product. It tries to extend the agent concept to be able to operate browsers. But unfortunately it is deemed to be a transitional product.

Clawdbot is a milestone in AI history. Not because of the technique. Its technique is quite shallow. Its product concept is groundbreaking. It has made clear to me that LLM is like CPU. Agents are software. LLM is hard to build. But is LLM that important? From the past years we can see: the technique is fast moving forward. Opensource LLMs today are better than the leading LLMs 1-2 years ago. In the next 1-2 years, the opensource LLMs will be equal if not better than Claude Opus4.6 and ChatGPT5.3 today. Opus4.6 takes huge effort to become the leading LLM in programming today. But soon some other LLMs will reach the similar performance of Opus4.6. Maybe Opus4.7 will be slightly advanced again. But the competition on LLMs level is brutle. To customers, there is no big difference. Just like today: if Intel's CPU is better, I will build an Interl CPU based PC. If tomorrow AMD publishes a better CPU, I can easily swith to AMD PC. At the moment, there is a Mac running M4 and a PC running Intel i7 on my desk. Switching between them is easy for me. Soon, LLM becomes replacable: I will switch to the leading LLM with low to no extra cost. What is matter to me is the software running on the machines. I cannot easily dump my blog website. I cannot easily get rid of my gmail. I cannot easily login to chrome with a new google account. Why? Because all the history of my user behavior and data are stored and bundled with the software. I have 10s of thousands of photos over 20 years time living in my google drive. Google drive reads my photos and learn my family and life. It can easily search all my kids' photos by name. I am not moving these photos to another cloud drive. Not as easy as buying a new PC. Agents will become more like the software. Agents are the ones that work closely with us, live with us, response to us, knows us and understand us. Agents do, not the underlying LLMs. The more we use the agents, the more they remember our preference, important setups, family members and memories. When I ask Clawdbot to send a message to my son's email, it is the Clawdbot who understnads me, not the underlying ChatGPT or Claude. I setup multiple LLMs under the hook and if one LLM runs out of quote, the other LLMs will kick in as backups. I set this up once and never care which LLM I am talking to. I am sure from time to time I have used different LLMs. But I don't care. The Clawdbot agent makes sure my requests are handled consistently. It is the Clawdbot I am interacting to, not the LLM. Clawdbot talks to LLMs, which clearly I care much less.

If you agree with the above, then a few things will become clear:

• Agents can live within our own machine. We can own the agents completely outright. All the agent related data belongs to us and can live in our own machine. Not like the Internet era where the big tech companies want to control all of our data. In the era of AI, this may be changed. Of course the LLMs are still run by the big tech companies. But LLMs are no longer that important. And if you still think LLMs are very important, think about how good those open source LLMs will become in 5 years time. The open source ones will be good enough to handle 99% of our daily requests. Good enough. Leave the state-of-art competition to the big tech companies. We can always set their LLMs as backups to only handle tough tasks.
• The user interface of software is foundmentally changed. Soon we don't need this or that kind of UI. UI helps us to manage information and data in the GUI era, where we need mouse, keyboard or finger to manage information and data. In the era of personal agents, we can simply UI into just an conversation based interface. Now it is the chat box. Future it maybe just mic and speaker. Frontend is dying. Backend can survive for a longer time.
• To best use agents, you need to teach it the desired workflow. Otherwise you cannot fully control the way it works. This is less important for works with clear process definition, such as solving a technical issue. To solve technical issue, you just tell clear what the issue is and the agents are pretty good to figure out a solution. But there are many things that don't have a standard right approach. If you want the agents to follow your approach, you need to teach it how. Once you teach it how, it can accomplish tasks in a much reliable way. This could mean to build many skills.
• There is much opportunity to explore in the best practice of managing memories, short-term or long-term. This can even become a research topic. I am not sure whether it will be very technical or not. But I feel like there is still a long way for us to come up with the best memory and context solution. Memory and context engineering is important. Clawdbot has some innovative approaches. I am sure some of its practice will be copied to wider industry soon.
• Integrate agents into communicating app is a product level innovation. It is a natrual fit. Unfortunately, those comm app companies didn't invite LLM in the first place. Otherwise I am sure the chatting would have now happening inside comm apps now.

I see Clawdbot as an innovation milestone in agent progressing. Many of its innovations are on product design level. I wouldn't be surprised if Clawdbot shapes the foundation of the future agent products.