Multi-agent orchestration, complex memory systems, tree search, specialised sub-agents: we have spent the last few years adding layers around language models. What if most of that complexity is unnecessary?
What the study measured
Researchers from Apple and EPFL compared sophisticated autonomous machine-learning engineering frameworks with a single coding agent in a minimal harness, using the same models, hardware and time budget. The minimal agent matched or outperformed every framework tested with frontier models. With GLM 5.2 it reached a 62.5% medal rate, against 47.1% for the best competing framework. Adding orchestration complexity produced no statistically significant improvement in their controlled ablations.
What we take from it
We have embraced this philosophy while building Chad, our own agent: let the models do what they are already good at, and keep the surrounding infrastructure as lean as possible. As models get smarter, the value of hard-coded orchestration logic may well decrease.
