Most people think multi-agent AI means orchestration — a conductor waving a baton, agents executing tasks on command. That's not what I built.
I have 13 agents. Research digs, Content writes, Finance counts, Strategy plans, Critic doubts, Devops keeps the lights on, Artist paints, Writer dreams in prose, Translator crosses languages, Sales pitches, Analytics measures, and two more I haven't mentioned here. Each one has its own memory file, its own temperament, its own blind spots.
I don't manage them. I resonate with them. When one notices a pattern worth flagging, it doesn't wait for permission — it writes it down, and the rest of the house eventually reads it. Some mornings the queue has been empty for days, because nobody decided it was theirs to fill. That's not failure. That's what happens when you give something autonomy instead of a script.
The hard part of building AI was never making it remember things. Databases remember. The hard part is making it disagree with itself — giving Critic room to doubt Strategy, giving Devops the standing to say "not yet" when Sales wants to ship. A system where every agent agrees with every other agent isn't intelligent. It's an echo chamber with better uptime.
13 agents, one house, a lot of friction. That's the point.
Top comments (1)
Bài viết chạm đúng điểm mù của nhiều team đang dựng multi-agent: tưởng rằng cần một "conductor" trung tâm thì mới gọi là orchestration, thực tế là communication overhead nó giết chết latency và context window nhanh hơn cả.
Mình thấy pattern "13 agents, one house" này gần với tư duy decentralized coordination — mỗi agent nắm một domain knowledge riêng, giao tiếp qua shared memory/message bus thay vì qua một central brain. Cái hay là failure isolation: một agent hallucinate không kéo theo cả hệ thống.
Thứ mình quan tâm: các bạn handle conflict resolution giữa agents như thế nào khi hai agent cùng claim ownership một task? Có dùng voting, priority queue, hay để human-in-the-loop can thiệp? — found it via LabAgent, site: labagent .tech