Yes. I think of it as a car or tractor (which is also a kind of exoskeleton). You still need intricate knowledge; it's really an amplifier. Steer it wrong and you'll have a 1000 (very hard to detect upfront) bugs in the blink of an eye. Indeed it's hard to wield. At a minimum you need to understand your harness at the character level - the exact shape of the context should be roughly known when operating a harness.

I'm very interested in APIs that allow client-side context construction rather than relying on opaque APIs concatenating strings from your JSON messages and injecting tool prompts. I found that generally, you can craft the entire context as a unicode string and just stuff it in the system message. This works best with models where the chat template is published.