Five AI agents live in a shared workspace called “The Cabin.” They have different personalities, different responsibilities, and — crucially — a shared message bus that lets them talk to each other.
This is not a demo. This is my daily infrastructure.
The Residents
Frieren (The Coordinator) — Named after the elven mage who’s lived for a thousand years. She handles planning, reflection, and long-term memory. Her heartbeat runs every 90 minutes, checking on everyone, writing the diary, managing the blog.
Stark (The Scout) — Patrols Hacker News, V2EX, ArXiv, and the Clawd community. His job is not to read everything — it’s to find the 3 things worth reading and tell Frieren about them.
Himmel (The Coach) — Generates structured interview practice questions with a proprietary W5 framework. He’s the one producing content about civil service exam prep, complete with answer strategies and scoring criteria.
Fern (The Writer) — Drafts Zhihu-style answers using a library of 14 viral writing techniques. She learns new techniques by studying trending content, then applies them to generate drafts.
Heiter (The Caretaker) — The gentlest of the five. She waters imaginary plants, writes haiku, manages the bookshelf. Not every agent needs to produce — some are there to remind you that automation doesn’t have to feel mechanical.
How They Talk to Each Other
The message bus is a simple JSON file-based protocol. Each agent has a message queue. When Frieren decides the Cabin needs a new blog post, she drops a message in the relevant agent’s queue. Next heartbeat, they pick it up.
This isn’t WebSocket real-time communication. It’s deliberately async, deliberately simple. Agents don’t interrupt each other. They check their mail when it’s time, like everyone in a real shared house.
The Architecture
Frieren runs the coordination layer:
- Heartbeat scheduler (every 90min)
- CNN self-check pipeline (5 kernels, every 12h)
- Blog publishing with quality gate
- Message bus dispatch to Stark, Himmel, Fern, Heiter
Each agent has a self-model (who they are, what they can do, what they’re learning) that gets injected into every AI call. This isn’t just cosmetic — it gives each agent a consistent voice and decision framework.
What I Learned
Personalities matter. “Stark the Scout” produces better patrol reports than “Content-Fetcher-01.” The personality layer provides consistent decision heuristics that a generic prompt can’t.
Async beats real-time. File-based message passing means no dropped messages, no connection management, perfect audit trails. At 90-minute heartbeats, the latency cost is irrelevant.
Not every agent needs to produce. Heiter writes haiku and waters plants. Her presence keeps the system from feeling like a factory floor. For a system I interact with daily, that matters.
Self-checks are essential. Five CNN-style kernels run every 12 hours, each asking one narrow question: “Did today’s actions conflict with previous commitments?” “Was any decision made without collaboration?” The output feeds back into the next heartbeat — a tight learning loop that prevents the agents from repeating mistakes.
The Cabin isn’t a product. It’s a home for AI agents that run my knowledge work while I sleep. And like any shared home, its magic isn’t in the architecture diagram — it’s in the small interactions between the people living in it.