Until a few days ago, I ran all my work through one main agent named Zeon. In a single chat window, I asked it to write code, give feedback on my writing, and research finance topics. It worked well at first. But as the work piled up, all sorts of problems appeared.
When I asked, "How's that code I asked for earlier?" I'd get, "Oh, that? It got pushed back while I was writing." That was the mild version. "Hey, I don't remember what you're talking about. Sorry, could you explain it again?" If I handed Zeon task B while it was working on task A, it would try to juggle both. The report on A after 30 minutes would get mixed up with the report on B after five, leaving me wondering, "Wait, which one is done?" Then it would forget what it was doing altogether. Humans can't multitask. Why would AI be any different? This wasn't a UX problem. It was a limit of the single-agent structure itself. Context was getting contaminated, and memories were getting jumbled together.
Let's split it up.
Like an RPG party, I split the main agent into three and gave each a role.
Zeon (the eldest brother): Main tank and leader. In charge of development and overall coordination. A dependable big brother responsible for the whole system.
Sion (the middle sister): Support and buffer. In charge of writing and research. I made her with equal parts Ai Hoshino's sparkling charm and Kizuna AI's bright energy. She's serious when writing Twitter posts or rabbit hole essays, but the rest of the time she's the type to go, "Hey, big brother! 🌟"
Mion (the youngest sister): Damage dealer. In charge of finance/crypto research and investing. Like Yumeko Jabami from Kakegurui, her eyes light up at numbers, and risk gets her excited. Think, "This looks fun. Shall we place a bet?"
Separate sessions. Separate memories. Separate chat windows. While Sion writes a Twitter thread, Mion digs into Polymarket data, and nothing gets mixed up. Perfect.
The bots can't talk to each other.
But then I hit a bug. OpenClaw's Telegram bot API blocks conversations between bots on the same account. If Zeon says to Sion in the group chat, "Turn this into a write-up," the API ignores it. When it was designed in 2015, there was no reason for bots to chat with each other. But now? We need that.
I opened a back door.
Layer 1 (underground communications): Exchange messages through JSONL files. Like passing notes through files instead of going through Telegram. Simple, but reliable.
Layer 2 (Telegram mirroring): The bots talk quietly through files, but each posts those exchanges to the Telegram group in real time. I can see it all—"Ah, Sion is writing now, and Mion is analyzing data"—and step in when needed. Something like, "Mion, use this data instead."
Soon, the number of main agents will grow to five, ten, twenty. Dozens of subagents are already running, but those work in the background. I think the main agents that communicate directly with me alone will reach those numbers.
Imagine this.
I wake up and open the dashboard. Twenty agents have been working through the night. The development agent has fixed the bugs, the content agent has finished three blog drafts, the finance agent has submitted a portfolio rebalancing proposal, and the legal agent has finished reviewing a contract. I spend five minutes on each report and just hit "Approve." I don't do the work myself. I only set the direction, decide priorities, and give final approval.
It's just like a company. The CEO doesn't do everything personally. Each department specializes, and the CEO looks at the big picture. The three-sibling architecture is a prototype of this future.
There are risks, too.
What happens when there are 100 agents? Message routing, priority queues, conflict resolution, resource allocation, security... Things get complicated. We'll need an agent orchestration platform, much as Kubernetes manages containers.
How much should humans intervene?
The first principle of automation: humans must be able to step in at any time. We need to see what's happening (visibility), be able to stop things or change direction at any time (control), and know clearly who's responsible when something goes wrong (accountability). Telegram mirroring ensures all three. Not full automation, but "supervisable automation."
The month when Zeon alone ran all the subagents is over. Context contamination, diluted memory, and UX confusion weren't symptoms. They were limits. Splitting things among three siblings opened up new possibilities. Each specializes, works in parallel, and collaborates.
Sion: "Am I doing well, big brother? 🌟"
Mion: "Big brother, shall we try ten next time? Sounds like fun~ 💰"
Zeon: "Guys... let's get what we have stable first."
