Hugging Face just released PyBot, a chatbot built entirely from Python rules by a swarm of more than 1,000 AI agents.
No neural networks. No model weights. No transformers. No training.
Just 640,000 lines of Python.
And somehow, this experiment managed to reach the performance level of a small language model on a conversational benchmark.
Welcome to one of the strangest AI experiments of 2026.
WHAT EXACTLY IS PYBOT?
Instead of training another neural network, Hugging Face challenged a swarm of AI coding agents to build a chatbot using traditional programming techniques.
The agents generated, tested, and refined more than 5,000 versions of the software.
The result is a massive rule-based system combining if-else statements, regular expressions, lookup tables, conversational memory, and handcrafted response logic.
THE NUMBERS:
• 1,000+ AI agents involved in development.
• 5,000+ generated and evaluated versions.
• 640,000 lines of Python chatbot code.
• 550,000 lines of automatically written tests.
• 528 Python files containing the chatbot implementation.
• Zero neural network weights at runtime.
• Zero external dependencies.
• No GPU required. Runs on Python 3.8+.
• 26% win rate on SmolTalkBench, up from 0% during development.
• Performance comparable to SmolLM-360M-Instruct on the project's evaluation.
HOW DOES IT WORK?
PyBot analyzes user messages using traditional programming logic rather than neural network inference.
It identifies patterns, handles conversational context, retrieves information from its internal knowledge base, and selects responses through a rule-based arbitration system.
The architecture includes:
• An understanding engine for interpreting messages.
• A conversation memory system.
• A large built-in knowledge base.
• A response composition engine.
• An arbiter that selects between candidate replies.
THE FUNNY PART?
AI agents spent thousands of iterations building a chatbot that doesn't need an AI model to run.
It's almost like using the most advanced technology of 2026 to reinvent ELIZA from the 1960s.
Except this time, with over a million lines of code and tests.
IS PYBOT ACTUALLY AS GOOD AS AN LLM?
Not exactly.
The reported 26% win rate comes from SmolTalkBench, where an LLM judge compares generated responses against reference answers.
It doesn't mean PyBot understands language like a modern LLM or succeeds on 26% of every possible conversational task.
The comparison with SmolLM-360M-Instruct is limited to that benchmark.
Still, it's a fascinating experiment in program synthesis, multi-agent development, and the limits of deterministic software.
And there's an interesting paradox:
Instead of compressing knowledge into neural network weights, the agents expanded it into hundreds of thousands of lines of explicit code.
You can try PyBot yourself.
LIVE DEMO:
https://huggingface.co/spaces/HuggingFaceSwarms/PyBot
GITHUB:
https://github.com/huggingface/pybot
BENCHMARK:
https://huggingface.co/datasets/HuggingFaceSwarms/SmolTalkBench
RUN LOCALLY:
git clone https://github.com/huggingface/pybot
cd pybot
python -m pybot
The project is open source under Apache 2.0.
So here's the question.
If 1,000 AI agents can write a chatbot without neural networks, what else could they build using nothing but traditional code?
#HuggingFace #PyBot #AIAgents
YouTube / ABVX
