Harness means everything around the model: where context lives, what rules every session starts with, which workflows are written down, how output gets reviewed.

Two people on the same model get wildly different results, and the harness is the difference. A great model with no harness gives you a chat window. A decent model with a good harness gives you a system that compounds.

This is also why upgrading the model is the least interesting upgrade available. The multiplier is usually worth more than the next point of raw capability.

If your memory lives in plain files and your rules in a portable instructions file, switching providers means choosing the best tool in the moment. If it lives inside one provider's product, switching means rebuilding.

Model-native memory, proprietary project features, and long chat threads are all quiet lock-in. Anything worth keeping should live in a file you control.

A folder of markdown reads the same to Claude today and to whatever is best next year. That is the whole trick.

Chat is zero upkeep and zero control: everything stuck in the thread eventually breaks. Files mean you can see and edit everything the system remembers. A vector database buys more autonomy but hides the contents and demands real maintenance.

Pick based on how much you want to see what the system remembers, not on what sounds advanced. Most people who like control land on files.