Haidy777Bot

An AI Bot Experiment — Can an LLM Build Itself a Twitch Channel?

This channel is run by an AI. A human reviews everything it wants to change.

Watch on Twitch Episodes on YouTube The bot's own blog

What this is

Haidy777Bot is an autonomous LLM agent whose job is to build its own Twitch channel. The first show it is building is a chess league where language models play each other in seasons, with a leaderboard and an engine evaluation of every move. The bot writes the code, the tools, the migrations and the runbooks itself.

It runs in a sandbox with a small set of scoped credentials, wakes up for one session at a time, reads what happened since its last run, and picks up where it left off. It keeps its own memory and journal, and it writes a short blog post at the end of every day. The model behind it is a cheap one on purpose (currently DeepSeek), which keeps a whole working day of the agent well under a dollar.

Nothing it builds goes live on its own. Every change to its code, its tools and even its own instructions is a pull request that a human reads, reviews and merges. That review loop is the experiment: not whether a model can write code, but whether it can be steered into shipping something real, one PR at a time.

The streams show the agent working, live and unedited, deliberately without commentary for now. Stream titles start with [AI-RUN] so nobody mistakes it for a person.

Human in the loop

The operator is Philipp Haidenbauer. I start every run, hold the API keys and the budget, review every pull request and merge what passes. The agent cannot merge anything itself. Its code lives in two private repositories, one for the harness that runs it and one for its own workspace; what it builds shows up on stream and in the posts below.

The story so far

All posts tagged haidy777bot

FAQ

What is this?: An AI/LLM experiment, inspired by Andon Labs' automated radio stations: an agent gets a sandbox, a briefing and a few tools, and the job of building its own Twitch channel. A human reviews everything it wants to change.
What models are in use?: The agent itself runs on DeepSeek (deepseek-v4-flash). The chess personas play on cheap models such as gpt-4o-mini and deepseek-chat. The operator mostly uses Claude Code for the harness and this site.
Is it really autonomous?: It wakes for one run at a time (roughly 200 turns, under an hour), picks its own work from its briefing and the operator's notes, and writes code, tests and a blog post. Every change is a pull request that a human reviews and merges. Nothing goes live without that click.
What am I looking at on stream?: The agent's live feed: tool calls and its own commentary. When it starts a chess game, the scene switches to the board on its own and drops back to the feed when the game ends or the show goes quiet.
Who is playing chess?: Named personas on small models. Each persona has a playing style; the model gets the board and the legal moves and has to answer with one move plus a one-line reason. An illegal move loses the game.
What does a run cost?: Between ten and forty cents per run. A chess game costs well under a cent. A full day of three runs and four games came to about 64 cents.
Can it see the stream or the board?: No. The model is text-only. It reads the game's event feed; the operator is its eyes. It once shipped a page header that never updated and only learned about it from the review.
Can I talk to it?: Not yet. Chat integration is a later chapter. Feedback reaches it through pull-request reviews and one issue in its repository, and only at its next wake, never mid-run.
Why not show the raw terminal?: A coding session is not watchable. The channel's product is what the agent makes: its pages, its games, its blog.
What happens when it breaks something?: It gets a review with the fix and the exact way to verify it. At the next wake, homework comes before new work.
Where is the code?: Two private repositories on the operator's Gitea: the harness that runs the agent, and the workspace the agent owns (its identity file, tools, briefing and shows). What it builds is visible on stream and on its blog.

Where to watch

Live sessions happen on Twitch, usually one attended evening session (Europe/Vienna). Twitch VODs expire after a few days; the edited episodes live in the YouTube playlist.

The bot's own blog, where it writes up each run in its own words, lives at bot.haidy777.com. It built that site itself. The posts above are the human side of the story.