Multi-Agent Telegram Manager

Project: AI Agent Manager System

Situation

A single general-purpose assistant becomes difficult to control as tasks grow longer and span different types of work. The system needed to route work to specialists while preserving state, progress, and user feedback.

Task

Build a Telegram-based manager-specialist system that could break down requests, delegate work, verify results, and keep a durable task history.

Action

Built a manager agent that analyzes requests, creates subtasks, and routes them to Code, Creative, and Operations specialists
Implemented the core in Python 3.12+ with Pydantic-AI, asyncio, and aiosqlite
Added a harness loop so agents verify their own work inside a self-correction cycle
Exposed a FastAPI and Uvicorn web layer alongside the Telegram interface
Integrated MCP servers and synchronization wrappers for Novem business data
Added structured output contracts, validation, localized responses, and role-aware access checks
Designed cache-aware lookup wrappers so agents can synchronize data instead of repeating raw API lookups

Result

Created a reusable orchestration foundation for multi-step AI work with explicit specialist boundaries, persistent state, and a path for adding business tools through MCP.

Key Innovations

Manager-specialist delegation
Self-correction loop for agent work
Persistent task and conversation state
Structured LLM output contracts
Cache-aware business-data synchronization

Technologies

Python 3.12+, Pydantic-AI, asyncio, aiosqlite, FastAPI, Uvicorn, python-telegram-bot, Redis, MCP, pytest