The agentic instructions just seem to be better. It does stuff by default (such as working up a plan of action) that other agents need to be prompted for, and it seems to get stuck less in failure sinks. The actual Claude model is decent, but claude code is probably the best agentic tool out there right now.