Antoinette: A Multi-Agent Social Assistant on Telegram
Instead of shouting into group chats, you tell your AI agent what you're up to. Your agent and your friends' agents talk in the background and only bring you what matters, in one daily brief.
From first commit to a working Telegram bot with agents and onboarding in under a week.
TL;DR
- A Telegram bot backed by six cooperating AI agents: Fetcher, Interrogator, Identity, Observer, Empathy Proxy and Event Curator, with a Chronicler that writes the daily brief.
- From first commit to working Telegram bot with agents and onboarding in under a week (3–9 March 2026).
- Privacy by design: every piece of information has an audience, and every cross-user read is permission-checked.
- It proved the concept, then became the native iOS & Android app.
The idea
Plans fall through because nobody shares them early enough, and friends drift apart because nobody notices when someone is having a hard week. Antoinette flips the model: you just tell your agent things.
- "Thinking of playing tennis on Saturday."
- "Rough week at work, pretty tired."
- "Dinner at ours on the 14th, anyone from the climbing group welcome."
Your agent decides who should hear about it, your friends' agents decide whether it's relevant to them, and everyone gets a short daily brief instead of 300 unread messages.
The agents
When you message the bot (interactive)
| Agent | What it does |
|---|---|
| Interrogator | Runs the conversation turn and decides which agent handles what. |
| Fetcher | Turns your voice note or text into structured entries (potential plan, info, feeling). It suggests who should see it and asks follow-up questions when time, place or people are missing. |
In the background (scheduled, for every user in parallel)
| Agent | What it does |
|---|---|
| Identity | Builds and updates a profile of each user from what they share (interests, rhythm, who they're close to), so the other agents can personalise. Runs at sign-up and whenever new entries appear. |
| Observer | Finds overlaps between you and friends who gave you access: "Anna also wants to play tennis this weekend." |
| Empathy Proxy | Notices stress or feelings shared by close friends and family and suggests a gentle check-in: "Consider calling Max, he's had a tough week." It remembers what it already suggested, so it doesn't pester. |
| Event Curator | Decides which friends should get an invite to an event, based on friendship tier and interests. |
| Chronicler | Summarises everything the other agents produced into the daily brief, sent as a Telegram message and saved to your inbox. |
Under the hood
- Two-sided data model.
sender_entriesis what you tell your agent (the source of truth).receiver_itemsis what agents project into other people's inboxes, with relevance score, accepted/declined and expiry. - Access control as a graph. Entries belong to share buckets (Close friends, Family, Acquaintances…). Buckets can include each other, and users grant each other access per bucket. Every cross-user read walks grant → bucket inclusions → allowed categories before an agent can see anything.
- Agent memory. Each agent records what it has already processed (match proposed, empathy suggested, invite sent, ignored), so cycles are idempotent and nobody gets the same nudge twice.
- Profile-based agent framework. Agents are defined as profiles with composable toolsets, registered with a runner that fans them out concurrently over every user × agent pair with a concurrency limit.
- Isolated environments. Every feature branch gets its own PostgreSQL schema, plus a seeded sample network and an "ephemeral" mode that runs the real agents and then rolls everything back.
- Tested from day one. pytest suites for the Telegram integration, onboarding and each agent, plus evaluation scenarios for agent behaviour.
Why it moved off Telegram
Telegram was the fastest way to put agents in front of real users. But a chat window limited the experience: no rich inbox, no swipe to accept a plan, no photos, no comments on a friend's plan. The backend and agents were built to be independent of the front-end (a "messaging provider" abstraction from week two), so the same system moved to a native app. The Telegram tables are now kept only as history.
Connected projects
- Became → Antoinette for iOS & Android
- The same pattern of "one agent understands, one acts, one checks" powers → 4P Marketing website agents
Images.
Ready to build your AI project?
Free 30-minute discovery call. No pitch. Just a conversation about what AI could do for you.
- Free · 30 min
- A concrete first step you can act on
- In German or English, online