The delivered architecture
An n8n workflow export with 39 functional nodes and 35 connections, including Telegram input, transcription, orchestration and four specialist groups.
Multi-agent automation / Workflow architecture
One conversation. A team of specialist agents.
A multi-agent assistant that brings email, scheduling, contacts and content into one Telegram conversation—with support for text and voice.
- Project
- Multi-Agent Operations
- Focus
- Multi-agent automation
- Technology
- n8n / Telegram / OpenAI / Gmail / Airtable
THE THINKING BEHIND THE BUILD
The challenge.
Requests rarely stay inside one tool. An assistant needs to interpret the request and delegate to the right capability while keeping the entry point simple.
The approach.
Telegram text and voice inputs route through a switch, with audio transcription where needed. An orchestrating assistant combines a model, memory and specialist workflow tools before responding through Telegram.
DELIVERED & AVAILABLE TO EXPLORE
What the build delivers.
The implementation
A central assistant delegates through specialist workflow tools for email, calendar, contacts and content. The diagram and node map are drawn from the exported source.
What you can inspect
Step through the illustrative calendar example below, open the complete node map or inspect the original workflow JSON on GitHub.
This is an inspectable workflow architecture. Connected execution still requires service credentials and specialist workflow references. A recorded live run and measured operating results are not available yet.
EXPLORE THE MULTI-AGENT SYSTEM
A request.
The right agent.
A useful reply.
Follow a calendar question from Telegram to a specialist and back. Each step explains one part of the architecture.
Illustrative walkthrough with scripted sample data. No live AI, calendar access or messages are sent.
01 / TELEGRAM INPUT
Start in one conversation.
YOU · EXAMPLE REQUEST
What’s on my calendar tomorrow?
A text message enters through Telegram. The source workflow also includes a separate path for voice transcription.
Telegram → Orchestrator02 / ORCHESTRATOR
Choose the right specialist.
REQUEST INTENT
Read the next day’s schedule.
EmailCalendar ↗ContactsContent
The central assistant has four specialist workflow tools. In this example, the calendar agent is the relevant route.
Orchestrator → Calendar agent03 / CALENDAR SPECIALIST
Look up the relevant events.
Team planning30 minutes · Sample event
Project review45 minutes · Sample event
The calendar specialist has an event lookup tool. These two events are sample data to show what the response could contain.
Calendar tool → Orchestrator04 / TELEGRAM RESPONSE
Bring the answer back.
ILLUSTRATIVE ASSISTANT REPLY
Tomorrow you have team planning at 09:30 and a project review at 14:00.
The assistant’s output returns to Telegram. This lookup example does not create or change calendar events.
Orchestrator → TelegramINSIDE THE PROJECT
Built around the details.
Email & calendar agents
Specialist tools cover email reading, drafting, replies and labels, alongside calendar lookup, event creation and updates.
Contacts & content agents
Airtable tools handle contact lookup and updates. A content specialist combines a language model with Tavily research.
Voice, memory & coordination
Voice messages pass through transcription. A central assistant combines conversational memory, specialist workflow tools, search and calculation.
FROM REQUEST TO RESPONSE
How the system
comes together.
- 01
Receive the request
Telegram receives a text or voice message. A switch routes voice through download and transcription, while text is prepared directly.
- 02
Interpret and coordinate
The central assistant combines the message with its model and conversation memory, and has access to specialist workflow tools.
- 03
Use the right specialist
Email, calendar, contact and content capabilities are exposed as separate workflow tools, with their own model and service connections.
- 04
Return to the conversation
The assistant’s output passes to a Telegram response node, keeping the user-facing interaction in one place.
A CLOSER LOOK AT THE BUILD
The complete workflow map.
All 39 nodes and 35 exported connections, redrawn from the source graph. Open the full-size diagram to inspect the model, memory and tool connections.
