LLMs drive today's chatbots and agents, but conversational programs are much older than LLMs. Sixty years ago, ELIZA held + conversations by matching patterns in what the user typed and answering with predefined templates.
+We wanted to revisit that idea at a much larger scale, so we gave a swarm of LLM agents the mission of building a chatbot out of + plain Python rules (if-else branches, regex patterns and lookup tables), with no model inside. The agents wrote and evaluated over + 5,000 versions of the bot, and its win rate on SmolTalkBench, where an LLM judge compares each reply with a reference answer, + climbed from almost 0% to 26%.
+The result is PyBot: about 430k lines of Python that answer everyday questions, plus 500k lines of tests the agents wrote to keep + it working. On SmolTalkBench it's on par with SmolLM-360M-Instruct (26.0% vs 26.6%), without a single neural network inside. You can + try it below, where you can see which of PyBot's modules produced the reply, and you can click any file to read the code behind the answer.
+