moziBot
A TypeScript multi-channel AI agent runtime: Telegram and Discord connected to one runtime with session persistence, task scheduling, job resumption, and structured observability.
Why I built it
AI agents are easy to demo and hard to operate. The interesting problems are the unglamorous ones: what happens to a scheduled job when the process dies, how sessions persist across channels, how you observe what an agent actually did. moziBot is my running experiment in treating an agent runtime like production infrastructure.
What it does
Connects messaging channels (Telegram, Discord) to a unified runtime with session management, task scheduling, reminders, job resumption, and agent job execution, plus a structured observability layer for watching it all work.
What it taught me
Agent reliability is a systems problem, not a prompting problem. Most of the code is scheduling, persistence, and recovery; the LLM call is the smallest part.