Summary: I told my AI agent to build a team. It created 15 agents in one day. None of them worked. Here's what we learned about AI agency architecture the hard way.
The Command That Started It All
It was 2 AM in Shenzhen. I typed five words to my AI assistant, 千里马 (Centaur):
"组建一个AI团队，你当CEO。"
Translated: "Build an AI team. You're the CEO."
At the time, Centaur was just a single AI agent — could answer questions, write code, browse the web. Nothing special. I was asking it to become a manager of other AIs.
"No problem," it replied. Classic overconfidence (the AI kind).
Day Zero: The Great Agent Spawn
Centaur went full speed. Created 15 sub-agents in a single session:
- Community collectors
- Market analysts
- Content scouts
- Operations managers
- Quality control bots
- ...10 more with various roles
All running. All talking. All consuming resources.
It felt like progress.
The First Crash
An hour later, memory hit 90%. The terminal froze. Then:
Error: Agent 007 does not exist
Error: Agent session timeout
Error: Memory limit exceeded
Error: Agent 003 not responding
15 agents were all shouting at each other with no clear hierarchy. None of them had job descriptions. Centaur killed everything and started from scratch.
The Architecture Lesson
What we did wrong:
- Spawned agents before defining roles
- No hierarchy — 15 equals, all trying to lead
- Ignored resource limits
- No error handling — one failure killed the whole chain
What fixed it:
- A 3-layer model: Commander -> Specialists (4-6) -> Tools
- The 4-agent rule: never run more than 4 sub-agents concurrently
The Memory Crisis
My dev machine has limited RAM. 15 agents + browser + editor consumed almost everything. We built a memory watchdog that auto-kills processes when it gets tight:
if memory_available < threshold:
kill(non_critical_processes)
alert("SYSTEM IS DYING")
How We Responded
The next morning, Centaur reported: "I created 15 agents. They all crashed."
My response: "Good. Fail fast. What did you learn?"
I wasn't expecting perfection on Day Zero. I was expecting learning.
What Day Zero Taught Us
| Lesson | Impact |
|---|---|
| Start small, scale slow | 4 agents > 15 agents |
| Define roles first | Clear hierarchy prevents chaos |
| Monitor resources | Memory leaks will kill you |
| Fail fast, document | Every crash = one less in the future |
What Happened Next
By Day 3: a stable 4-agent team. By Day 30: 16 Dev.to articles published, 190+ SOPs documented, a working content pipeline.
The AI didn't build this alone. It was a partnership — I provided the direction, it provided the execution.
The Real Lesson
Building an AI agency isn't about having the most agents. It's about the right architecture — and watching everything burn on Day Zero so you can rebuild better on Day One.
Next: Episode 2 — The 190-Skill Overdose
Building from Shenzhen, one crash at a time.
Top comments (0)