n0rds3c.dev / notes / draft
Why most agents feel like demos, not products
. Draft. Not finished; the argument is here, the polish is not.
Most agent demos are honest. They show a real model doing a real task. What they leave out is everyone who is not the person holding the mouse: the second customer whose data must never appear, the reviewer who needs to know what happened last Tuesday, and the integration that calls the API directly and never sees the guardrail you put in the interface.
That gap is the whole distance between a demo and a product. It is not a model problem and no model release closes it. It is tenant boundaries on every data path, approvals on every consequential action, a trail you can verify, and a failure story for the day the model is wrong. Each of those is unglamorous and each is the reason a team can keep using the thing after the first bad week.
The tell is what happens when you ask to see the log. A demo has a transcript. A product has a record with an approver, a tenant, and a hash of the row before. If the answer to “can I audit this” is a shrug, you are looking at a demo, however good the model is.
I am not against demos. I am against shipping them. Build the boring parts first and the demo gets better, because now it is a demo of something that will still be there next quarter.