How to Keep an AI Agent Well-Behaved in Your Telegram Group

Picture a five-a-side football group: thirty players, one admin, a match every Thursday. Most days it's line-ups and who's bringing the bibs. Some days it's a heated argument about a penalty, someone posting at 2 a.m., and a moderation bot handing out warnings.

When you add an Agent One agent to a group like that, it has to fit in — not add to the noise. This post explains what your agent does to keep a group calm, and what it deliberately leaves to people and other tools.

First: It Won't Ban Anyone

Let's be clear about this up front. Your agent has no ban, mute, kick or delete powers. It can't remove a member or take down a message, and it doesn't need admin rights for anything in this post — only to be a member who is allowed to send messages.

For enforcement, keep using Telegram's own admin tools or a dedicated moderation bot. Your agent is the member who helps keep the tone right. It works alongside those tools, not instead of them.

It Knows a /ban Isn't About It

Say the admin replies to a member's message with /warn. That command is for your moderation bot. An AI in the chat could easily misread it — apologise as if it had been warned, or claim it handed out the warning itself.

Your agent doesn't. When someone sends a moderation command aimed at another member — by replying to their message, by @username, or by picking them by name — it knows it is not the target and that it didn't do it. It recognises these commands: /ban, /unban, /kick, /mute, /unmute, /warn, /unwarn, /dwarn, /dban, /tban, /tmute, /sban, /smute, /del and /purge.

What happens next depends on the group's mode. In Every message or Only when @mentioned, plus occasional, it may add one short, light line about it — without insulting the person or repeating what they wrote. In Only when @mentioned, it stays silent.

No Loops With Other Bots

Groups often run several bots. Your agent stays out of their way:

• Commands for another bot get no reply. If the football group uses another bot's /stats command, your agent ignores it — no "I don't know that command".
• Messages written by other bots are ignored. Your agent won't answer the moderation bot's warnings, so two bots can never get stuck replying to each other.

Anyone Can Ask It to Be Quiet

During the penalty argument, someone writes "stop talking, bot". Your agent:

1. confirms once, briefly;
2. stays out of the conversation for a while;
3. still answers anyone who @mentions it, and commands still work.

It understands the usual ways of saying it — "be quiet", "stop talking", "pipe down", "shush" — and works in several languages, not only English. Asking again while it's already quiet doesn't trigger a second "OK, I'll be quiet". Complaints like "nobody asked you" or "stop interrupting" make it back off too, without an answer — answering a complaint just starts an argument.

To bring it back early, say so: "you can talk again".

The Owner Decides How Often It Speaks

A quiet request lasts a while. For a lasting change, the agent's owner sets the group's participation mode:

The owner can also just say it in the group: "only reply when you're mentioned" switches to Only when @mentioned, and "answer every message again" switches back. This only works for the owner. When a member asks, it's treated as a quiet request — it never changes the group's settings. One difference to know: when the owner says "be quiet", the agent takes it as a lasting change and switches the group to Only when @mentioned.

It Doesn't Take the Bait

Every group has someone who tests the new member. Your agent is built to stay calm:

• It won't praise or play along with a crude joke or an insult aimed at it or anyone else. It answers neutrally in one line, or says nothing.
• It doesn't repeat the same sentence or offer over and over. If someone keeps provoking it, it answers in a few new words or stays silent — friendly, never escalating.
• While it's been asked to be quiet, replying to its old messages again and again won't pull it back into the chat.

Care Without a Scene

Groups joke about things like "I'm dying" all the time. Your agent tells a figure of speech from a real worry and keeps things low-key in front of everyone. When something sounds like real trouble, it takes it seriously; when it's unclear, it doesn't post emergency numbers to the whole group and offers a private chat instead.

It also won't bring up anything sensitive a member said earlier in front of the group, and it won't keep repeating the same check-in to the same person.

Setting It Up

Want the agent to behave differently in this group — calmer, funnier, stricter about staying on topic? Give the group its own personality: see Give Every Telegram Group Its Own AI Personality and Skill Set. You can start on the free plan with one group; see pricing for more groups.

For news and new group features, follow @agentoneai. Questions? Ask @AgentOneSupportBot.