I find it difficult to distinguish agent coding from just a code harness. They refer to different things, but those agents dont exist without a harness that manages them.

The harness though is real; the LLM does not do anything interesting without a set of tools, specific prompts to use them, and taking a lot of forced context planning out of the dev's hands.

But yeah, usually when you mix business interests without customer concerns, those business interests develop dark patterns so they can invent concerns and then solve them.