I handed development over to AI agents: the story of a self-evolving Telegram bot

A developer decided to fully automate the development of his Telegram bot using AI agents: at night, agents add new features and improve code, in the morning he simply reads the list of changes and enjoys the results.

How It All Started

It all began in 2022, when the author moved to Georgia and decided to learn the local language. Duolingo didn't support Georgian, but offered Klingon instead. So he decided to make his own version — a Telegram bot for daily learning. At first, it was something simple: save a word, view a translation, take a quiz with new words. This was enough, and the bot developed slowly. But then an idea came: what if I hand over the entire development process to AI and just watch the bot grow?

Overnight Development with LLM Agents

The plan was simple and tempting: enable modern language models that would work at night. They add new features, improve code, fix bugs. In the morning, all that's left is to read the changelog and see what turned out. The idea seemed perfect for automating routine work:

• Agents can write simple code without human help
• At least one small feature is added every day
• Days of manual coding are saved
• Results are visible right after a few hours of work

It sounds like a self-developing startup — a project that grows on its own while you sleep. But in reality, everything is more complicated.

What Actually Happened

AI agents really do write code and add features. But the problem is that development is not simply a sequence of independent tasks. A new feature can break an old one. Architecture requires thoughtful decisions. Priorities change depending on how users interact with the bot. Agents work well on individual components: write a search function, refactor code, add a test. But a holistic vision of the project, a development strategy, choosing between different approaches — this requires a human. Without intermediate checks and feedback, overnight automation can lead to uncontrolled code sprawl.

What This Means

The future of development will be hybrid. AI agents are excellent helpers for speeding up routine work: tests, refactoring, documentation, simple features. But strategic decisions, architectural conclusions, and priority management remain with humans. The dream of fully automated development runs up against the complexity of real-world projects.