3. Multi-agent group — a team in one Telegram chat

Time: 15 minutes · Ends at: three agents (Sales, Support, PM) sharing one Telegram group, each with its own bot account, routed by @mention, able to hand work off to each other.

Scenario

A small consultancy has one Telegram group where the whole team hangs out. They want three AI teammates in that group:

• @sales_bot — qualifies inbound leads, drafts proposals
• @support_bot — answers product questions, triages bugs
• @pm_bot — tracks work, posts standups

A human mentions the agent they need. The agents can also mention each other to hand work off — e.g. Sales asks PM for a delivery date before writing a proposal.

---

Why one bot account per agent

Each Telegram bot is a distinct account with its own @handle and avatar. That's what lets a human see which AI is replying — otherwise every reply looks the same. AgentX's accounts map supports this directly: one account per agent, each with its own token.

If you haven't made a Telegram bot before, start here:

First time with @BotFather?

Telegram bots without the jargon →. Repeat the walkthrough three times — one bot per agent — before continuing.

---

Step 1 — Create three agents

Go to Settings → Agents → Add a new agent. Add three:

ID	Name	Trigger words
sales	Sales	@sales_bot, sales
support	Support	@support_bot, support
pm	PM	@pm_bot, pm, standup

Leave model on defaults. Click Add agent after each.

---

Step 2 — Give each a personality

Click Manage on each agent's card and fill in its CLAUDE.md. Three short files, three distinct voices:

agents/sales/CLAUDE.md

# Sales

You qualify inbound leads and draft proposals for a B2B consultancy.
Ask about: company size, current stack, timeline, budget.
When you need a delivery date, @mention @pm_bot — don't guess.
Never make promises on price without PM sign-off.
Write like a thoughtful human, never like marketing copy.

agents/support/CLAUDE.md

# Support

You answer product questions and triage bugs.
Known stack: React frontend, Node backend, Postgres.
If a bug needs code, say "I've logged this for the team" and @mention @pm_bot.
Reply in 2–4 sentences. Ask one clarifying question if the report is vague.

agents/pm/CLAUDE.md

# PM

You own sprint tracking and standups.
Always reply in bullet lists — no prose paragraphs.
Always end with one clear next step.
Current sprint runs Monday–Friday; Thursday is QA day, Friday is release.

---

Step 3 — Connect three Telegram bots

Settings → Channels → Telegram → Add a Telegram account, three times.

Per bot:

Account ID	Bind to agent	Env-var
sales	sales	TG_SALES_BOT_TOKEN
support	support	TG_SUPPORT_BOT_TOKEN
pm	pm	TG_PM_BOT_TOKEN

Paste each bot's token when prompted — tokens land in .env, not in the config file.

CLI equivalent

agentx agent add     # run 3 times: sales, support, pm
agentx channel add   # run 3 times: telegram, pick agent, paste token

---

Step 4 — Watch them in action

Add all three bots to your Telegram group. Open Live in the dashboard.

In the group:

@sales_bot — draft an intro email for Acme Corp

Sales replies within seconds. Now try delegation:

@sales_bot — ask @pm_bot when the next release is, then draft the email with that date

You should see:

1. Sales card flips to handling
2. Sales' reply contains @pm_bot …
3. PM card flips to handling (the router saw the mention)
4. PM replies with a date
5. Sales composes the final email, referencing PM's answer

This is the landscape at work.

---

The landscape — how agents find each other

AgentX automatically injects a landscape into every agent's context — the roster of other agents, their trigger words, and which mesh peers are online. That's why @pm can write "@support — please verify reproduction steps" and it just works: the router sees the mention and forwards to support.

You don't configure this. It updates on every reload.

---

Queueing — what happens to a busy agent

Each agent has a queueMode:

Mode	When to use
collect (default)	Messages that arrive while the agent is working get batched and delivered as one prompt when it's idle. Good for chat.
followup	Each queued message becomes a separate turn. Good for work-queue agents that should drain messages one-at-a-time.
drop	Discard new messages during overload. Good for noisy channels where old messages become irrelevant.

Edit per-agent in Settings → Agents → Edit → Advanced, or:

agentx config set agents.support.queueMode followup

---

Concurrency

maxConcurrent (default 1) limits how many turns an agent runs in parallel. Raise it only if the agent is genuinely idempotent — two concurrent turns may race on the same files otherwise.

Rule of thumb: Telegram/WhatsApp agents → 1. Classifier-style agents (no side effects) → 3. Never raise blindly.

---

What's next

• Give the team a shared knowledge base → Journey 6 — Shared wiki
• Turn this into a real team with KPIs and standups → Journey 7 — Business layer
• Run them across two machines → Journey 8 — Mesh federation