All field notes

AI · Retrospective

Razortooth AI: an agent in the community chat

A small white robot among speech bubbles, one outlined in blue.
Concept illustration · AI-generated

On 3 December our Telegram channel opened the month with a preview of Razortooth AI. On 9 December we started a meme contest around it: make a Razortooth AI meme, post it on X with #RazortoothMeme, and the meme with the most hearts by 16 December wins.

Razortooth AI is an in-house agent for our own Telegram community. Its two jobs are taking part in the community chat and monitoring the group. The build is the team's; I oversee it as CEO. This is how I think about what an agent like this is for, and where its limits belong.

What it is there for

A game community on Telegram does not keep office hours. Questions arrive at any time, and many of them are questions that have been answered before: how a feature works, where a guide is, what an announcement meant. A team of people cannot be present for all of it. An agent can be, and it can point people to what the studio has already written.

The second job is quieter. Crypto communities attract people who pretend to be admins, links that imitate a project's own, and messages that should not stay up for long. Watching a busy group is constant, repetitive reading. That is work where a tireless reader helps the people moderating, rather than replacing them.

What it should not be trusted to say

The more useful design question is not what an agent in a chat can say but what it must not. Underneath everything I would put one plain rule: an agent like this should hold no keys and move no funds, so nothing it says can become a transaction. Above that rule, these are the lines I care about for any agent that speaks in a community's name:

  • It is plainly an agent. Members should always be able to tell it apart from a person on the team, and it should never sign off as one.
  • It does not predict prices or tell anyone what to buy. A game community's chat is not the place for advice, from a person or a program.
  • It does not make promises for the studio. Dates, rewards and rules come from announcements written and checked by people on the team. An agent that improvises a launch date has created a commitment nobody agreed to.
  • It does not settle disputes between members. It can point to the published rules. Judging who was in the wrong is a job for a moderator who can be asked to explain the decision.
  • It says when it does not have an answer. Handing a question to a person is a good outcome, not a failure.

Any agent that writes its own replies can produce a confident wrong answer. In a chat where some questions touch money, the boundary has to come from what the agent is able to do, not only from how it has been asked to behave. If a wrong answer cannot move anything, it is an error to correct. If it can, it is an incident.

Not an agent with a token

This season has brought a wave of AI agents launched alongside their own tokens, where the agent's posts and the token's chart are promoted together. Those projects are measured by attention.

An agent in a community chat should be measured on duller things. Do members get correct answers? Are scam messages noticed before someone clicks? Is the moderators' day easier? Does a person step in quickly when a question is beyond what the agent should handle? None of these appears on a chart, and all of them decide whether the community trusts the space it gathers in. Our December preview described Razortooth AI in terms of the community, and that is the frame I want people to judge it in.

Why start with a meme contest

Introducing an agent through a meme contest can look like a marketing detail. It does something useful. Members meet Razortooth as a character before they meet it as a voice in the chat, and they learn its name from each other rather than from an announcement. A community that already treats an agent as a character, not an oracle, is better placed to push back when it gets something wrong.

Where it sits

The way I picture Razortooth AI is as part of how the team runs the chat, not as a replacement for the team. People write the announcements, make the rules and take the decisions that affect someone's account or place in the group. The agent carries the repetitive load around them: the repeated question, the suspicious link, the message at an hour when nobody else is awake.

That division is not a limit on what the technology might do. It is a choice about who answers for what happens in a community, and I want that answer to stay with people.