From language model to agent, in one guide

Hernan Huwyler's guide on how language models become agents is written for cloud architects. As far as I can tell, it never mentions the person who has to trust the output.

The line I'd keep: agents only work where somebody can check the answer. The guide says repeated sampling "only works with a reliable verifier." Code has tests. Most of what designers make doesn't. So when making things is cheap, checking is the expensive part. That's my read, not the guide's.

Checking is an interface problem. A diff is an interface. Two things side by side is an interface. A confident paragraph saying "I did the right thing" is not. Show me the evidence next to the claim.

Also, the guide is part of a certification program the author is launching. I can't prove that shapes the writing, but I noticed.