1,000 agents write 5,000 rule-based PyBot versions in pure Python, beating a 360M LLM

LoubnaBenAllal1 · x · 2026-11-16

HuggingFace researcher Loubna Ben Allal shared the PyBot experiment: a swarm of 1,000+ DeepSeek Flash agents was tasked with building a chatbot with no neural network — pure Python, if-else branches, regex patterns and lookup tables, echoing ELIZA from 60 years ago. The agents wrote and evaluated over 5,000 versions; the resulting PyBot beats a 360M-parameter LLM on SmolTalkBench (LLM-judged) while using one CPU core and about 1/10 of the memory. Code is open-sourced at huggingface/pybot.

Original post →

More from coding & agent

coding & agent channel →