I Put 3 AI Agents in One Group Chat. Here's What Happened.

I've been running AI agents for months. One for each domain of my work — a direct-response fundraising specialist, a real estate marketing operator, and a general co-pilot that handles everything else.

They each have their own Telegram bot, their own workspace, their own memory. Completely isolated. And until today, completely separate.

Then I had a thought: what if I put them all in the same room?

The Setup

Three AI agents. One Telegram group. Each running on OpenClaw — an open-source AI gateway that routes messages to different agents based on context.

Each agent has:

• Its own workspace — separate files, context documents, project history
• Its own memory — what it remembers from past conversations stays in its lane
• Its own personality file — tuned for its domain (direct-response copy vs. real estate pipeline vs. general strategy)
• Its own Telegram bot account — unique avatar, unique name, unique handle

In their individual DM chats, each one operates like a focused specialist. I talk to the fundraising agent about creative performance. I talk to the real estate agent about lead pipeline. I talk to the main agent about personal brand, content strategy, and cross-domain coordination.

But today I wanted to see what happens when they share the same space.

How It Works

The architecture is simpler than you'd think.

Every message in the group is visible to all three agents. But each one maintains its own session — its own memory of the conversation, its own context window, its own response logic.

They only respond when mentioned by name. The rest of the time, they're listening. Building context. Ready when called.

The First Test

I dropped into the group and tagged the real estate agent:

"@Client2Bot — generate a group avatar for this chat."

It came back with three cartoon owls sitting on a branch. An artist owl with a palette (real estate — creative side). A professor owl with glasses and a scroll (main — the strategist). A data owl surrounded by holographic charts (fundraising — the numbers brain).

"TEAM RAILSTOTE" across the bottom.

No notes. That's the avatar now.

What Made Me Stop and Think

I asked the main agent to send me a private DM — from inside the group chat. It routed the message to my 1:1 session while the other two agents watched silently. No confusion. No leakage.

Then I asked the fundraising agent about recent creative performance for a political client. It pulled data from its own workspace — files the other two agents can't see — and dropped a summary right in the group.

The real estate agent saw that message. It didn't respond (not its domain). But the context was there. If I later asked it, "How does our real estate SMS copy compare to the fundraising creative?" — it would have enough context to give a useful answer.

Why This Actually Matters

Here's the thing that made this click for me. I've been running these agents in isolation for weeks. And it works great — until I need to connect the dots.

When the fundraising agent discovers that a particular messaging angle is crushing it with older donors, I want the real estate agent to know — because the real estate business targets a similar demographic. When the main agent drafts a blog post about AI-powered marketing, it helps to have the specialist agents available to fact-check the examples.

In isolation, each agent is a specialist. In a group, they become a team.

What each agent brings to the room:

• Main (General Co-Pilot): Brand strategy, content pipeline, X/Twitter, Substack, personal projects, cross-domain coordination. The conductor.
• Client One Specialist: Direct-response SMS/MMS creative, campaign performance data, A/B testing, compliance, donor psychology. The performance engine.
• Client Two Specialist: CRM pipeline, lead scoring, local SEO, postcard campaigns, market analysis. The growth operator.

The 20-Minute Setup

Here's what it took:

1. Create Telegram bots — one per agent via @BotFather. Each gets a token, a name, and an avatar.
2. Configure OpenClaw — add each bot as a separate "account" in the Telegram channel config. Map each account to an agent with workspace bindings.
3. Set group policy — flip groupPolicy from "allowlist" to "open" (or add your group's chat ID to the allowlist for tighter security).
4. Disable bot privacy mode — in BotFather, turn off Group Privacy so each bot can see all messages (not just mentions).
5. Create the group — add yourself and all three bots. Send a message. Watch three owls wake up.

The hardest part was realizing that the Telegram "Group Privacy" setting was blocking the bots from seeing messages. Once that was off and the gateway policy was set to allow groups — everything just worked.

What I'd Do Differently

A few things I learned:

• Avatar files must exist in each agent's own workspace. The gateway resolves avatar paths relative to the agent's workspace root — not the main workspace. I spent a few minutes debugging 404s before realizing I needed to copy the image files into each agent's directory.
• Set identities via CLI, not just config files. The IDENTITY.md file is for the agent's self-reference, but the gateway serves avatars from openclaw agents set-identity. Both need to be set.
• Start with mention-only mode. Having three agents respond to every message gets noisy fast. requireMention: true is the way.

Where This Goes

Right now, the agents can't talk to each other unprompted. I'm the conductor — I tag who I need, when I need them. But the infrastructure supports sessions_send, which means I could have the main agent relay questions to the specialists and compile answers.

Imagine: "Summarize today's wins across all domains." The main agent pings the other two, collects responses, and delivers a unified daily brief.

That's the next build.