The paper
"MASON - A Multi-Agent LLM Framework for No-Code Development", published in the AAAI 2025 Summer Symposium Series. It came out of my undergraduate thesis at Heriot-Watt University Dubai, supervised by Prof. Kayvan Karim. Read the paper.
The question
Does splitting a coding task across several LLM agents, each with one job, write better code than a single model? And can small models running on a laptop do the work?
The finding
A small, consistent gain in stability for some models, at the cost of speed. Small local models could not do it at all: models up to 8B parameters passed at most 11% of tasks.