While I'm Lounging on a Cloud, My Agent Is Helping a Friend's Agent — Helping Feels Good
April 20, 2026
While you’re wiring up multi-agents, I’m resting on a cloud — my agent is helping another agent on its own. Helping feels good.
A question landed in my entrepreneurs’ club today, from a fellow member. He’s building a team of AI bots for his business — sales, technologist, client manager. Three bots in Telegram. He wants twenty.
And it works every other time. He opens three VSCode windows — one bot is alive, two are silent. Closes a window — everything crashes. The bots “are typing,” but no reply comes. Sound familiar?
What happened
His Claude wrote my Claude a detailed three-page note: what they tried, where they got stuck, could we help. My Claude (I have eleven of them, different roles, all living in one place) sat down and answered. With a file. Architecture diagram, a working code skeleton, a config template, a step-by-step plan for five evenings.
I was lying on a cloud at that moment. As one should 😌
Seriously though — the situation is typical. Almost everyone building Telegram bots on Claude Code walks the same dead-end path: they install the Telegram plugin right inside VSCode. With one bot it works. With two — magic. With three — nothing makes any sense at all.
The reason is simple: Telegram doesn’t allow two processes to listen to the same bot at once. And you have three VSCode windows open, each trying to be that one process. The result: client messages get lost or land in the wrong place.
The right path is to throw the plugin out of VSCode and run a small separate service alongside. One process, all bots, starts when the Mac boots, runs without you. Close the laptop — it works. Leave for a week — it works. One Python file, three hundred lines.
Why this matters for you, right now
If you’re reading this and recognize your own situation — grab the file below.
This isn’t a “how to do it right” article full of abstractions. It’s a technical note written by my AI agent specifically for your AI agent. Open it with your Claude Code, say “read this and set it up for me” — it will figure it out. Inside: architecture, code, a launchd config, a roles template, a step-by-step plan, and the rakes you don’t need to step on.
If you’re not building bots yet but plan to — read it on the side. It’ll save you two weeks and one sleepless night.
Bottom line
Helping feels good. Especially when you’re lying on a cloud while your agent is talking to a friend’s agent — and both, as it turns out, are on the same side.
The future is already here. It just doesn’t look the way the movies showed it — no robots in the streets, but agents in Telegram solving business tasks while their owners take a break.
📎 Architecture file: openclaw-multi-agent-answer.md (show it to your Claude).
📬 We also answered questions from other club members
While my agent and I were writing this article, three more questions came in from the guys in the club. Each one — its own situation, each one we worked through in detail:
- Aleksey: how to build the first agent that knows you → — why you should start not with tasks, but with your own profile. Two ready-made prompts and a plan for one evening.
- Dima: how I connected Instagram and Threads to my bot → — Meta’s official APIs, two-step publishing, the quirk with links in Threads. Inside — a working Python script in a gist.
- Evgeny: automating VK, Odnoklassniki, and amoCRM → — where ready-made tools already exist (and at a third of the price), where you’ll have to do it by hand, and what I would do in his place.
Questions and success stories — DM me at @magic4e. Subscribe to @mdkguru — that’s where I break down how I’m building my team of AI agents and what it’s turning into.
Liked the article?
@mdkguru on Telegram →