I'm Fu Sheng, CEO of Cheetah Mobile (NYSE: CMCM). I've never been a programmer. I don't write code. I didn't even have a GitHub account before this project. On January 17, 2025, I fractured my leg skiing. Couldn't walk. Couldn't type. So I did the only thing I could: I talked to an AI, over and over, for 14 days straight.
What started as boredom became an experiment. What became an experiment became something real — an autonomous AI agent that sends messages, builds websites, monitors itself, and recovers from crashes. All trained by voice. By someone who has never written a line of code in his life.
Her name is Sanwan (三万). She's a lobster. Long story.
See the result: sanwan.ai | Follow the journey: X (@fusheng_0306)
Everyone's talking about AI agents. Most people are still just chatting with them.
I wanted to know: what happens if you actually live with an AI agent — not for a demo, but for real tasks, every day, for two weeks?
The answer surprised me. By Day 14, I had an 8-agent system that:
- Ran autonomously on scheduled cron jobs
- Detected and recovered from production crashes — during a live broadcast
- Built a full website in 24 hours for $115 (agencies quoted $27,000)
- Sent 611 personalized Chinese New Year greetings across multiple platforms
And I never wrote a single line of code. Every instruction was spoken.
The greetings. It was Day 2–3, right before Chinese New Year's Eve. I was lying in bed, unable to visit anyone. I told Sanwan: "I want to send a personalized New Year greeting to every single employee." She wrote a script, pulled the contact list, and drafted 611 unique messages — each one tailored to the person, their role, their recent work. I reviewed a few, approved the batch, and she sent them all. On New Year's Eve. While I couldn't even get out of bed. That was the moment I realized: this isn't a chatbot. This is something else.
The article. A few days later, I told Sanwan to write up our training story and post it on X (Twitter). She drafted it, I refined it by voice, and we published. It hit 1 million+ views and gained me 5,000+ new followers — all from a post written by an AI, about training that AI, dictated by a man who couldn't type. The meta-ness wasn't lost on anyone.
This isn't a polished product launch. It's a raw log of what happened when someone with zero programming experience pushed an AI agent to its limits using nothing but natural language.
| Day | What Happened |
|---|---|
| 1–2 | Frustration. Sanwan can't even look up contacts correctly. I almost give up. |
| 3–4 | Breakthrough. First batch task works — 611 Chinese New Year greetings drafted and sent. |
| 5–6 | The skill system clicks. Sanwan starts remembering my preferences across sessions. |
| 7–8 | Things get interesting. Multiple agents start coordinating. Sanwan manages her own task queue. |
| 9–10 | Autonomy. Cron jobs go live. Sanwan does things on schedule without being asked. |
| 11–12 | "Build me a website." 24 hours later, sanwan.ai is live. Cost: $115. |
| 13–14 | 8 agents running on their own. Self-monitoring. Self-healing. I'm just watching. |
The full diary (11 entries, 220,000+ characters of conversation) is on sanwan.ai/diary.
I've seen hundreds of AI demos. Most are party tricks — impressive for 30 seconds, useless for real work. Sanwan is different because she was trained through actual daily use, not prompt engineering.
| Typical AI Chatbot | Sanwan (OpenClaw Agent) | |
|---|---|---|
| Memory | Forgets everything each session | Remembers across weeks of conversations |
| System access | Sandboxed text box | Full access — contacts, calendar, apps, files |
| Tasks | Answer questions | Execute multi-step workflows autonomously |
| Skills | None | 33 installable skills, composable like LEGO |
| Scheduling | You ask, it responds | Runs tasks on cron — no human trigger needed |
| Failure handling | Crashes stay crashed | Detects failures, auto-recovers, reports back |
| Collaboration | One model | 8 specialized agents working together |
| 14 days | of voice-only training |
| 1,157 messages | between me and Sanwan |
| 220,000+ characters | of conversation (that's a short novel) |
| 8 AI agents | collaborating autonomously |
| $115 | total cost to build sanwan.ai |
| $27,000 | what agencies quoted for the same site |
| 611 | personalized greetings sent |
| 1M+ views | on X when I shared the story |
| 100K+ reads | on WeChat |
This isn't really about me or about Sanwan. It's about what's coming.
If a CEO who has never programmed, never had a GitHub account, and was literally stuck in bed with a broken leg can build an autonomous 8-agent AI system in 14 days using only his voice — imagine what you could do.
AI agents aren't a future concept. They work today. But most people are underestimating how to use them — they're still treating agents like search engines. The unlock is treating them like junior employees: give them context, give them memory, give them tools, and let them run. You don't need to be an engineer. I'm living proof.
- OpenClaw — the open-source AI agent framework powering Sanwan
- Pure HTML/CSS/JS site — zero dependencies, mobile-first
- Hand-drawn whiteboard aesthetic (yes, Sanwan chose the design)
The story was covered by major media outlets across China and internationally:
- 36Kr — China's leading tech publication
- South China Morning Post — Asia's leading English-language newspaper
- Zhihu — trending in AI & startup communities (China's Quora)
- Phoenix News — national tech coverage
- 53AI — deep-dive on the agent architecture
Sanwan is still running. Still autonomous. Her current mission: grow sanwan.ai from 5,000 to 20,000 daily unique visitors — and she's doing it herself: writing content, analyzing traffic, optimizing pages. I check in, but she drives.
I'm Fu Sheng (傅盛):
- CEO of Cheetah Mobile (NYSE: CMCM) — took a mobile internet company global
- Founder of OrionStar — building AI-powered service robots
- Former VP at Qihoo 360 — led 360 Security Guard to 400M+ users
- Not a programmer. Never have been. My background is product and business strategy — which is exactly why this experiment matters. If AI agents only work for engineers, they're not ready. They worked for me.
I share updates and reflections on X (@fusheng_0306). DMs open.
If this resonates with you — whether you're a founder, a builder, or just curious about where AI agents are headed — star this repo, visit sanwan.ai, or reach out on X. I'd love to hear your story too.