A weekend demo and a production agent
The demo will work. Here is what the six months after it are for.
Every founder we talk to has the same first reaction, and it is the right one: "we could build this in a weekend." You could. A model, a few tools from your API, a Saturday, and you have an agent that answers questions about your product and reschedules an appointment on stage.
We are not going to pretend it won't work. It will. What follows is the list of things the weekend does not get you, written as plainly as we can, because a technical buyer who is told the demo is impossible stops listening.
A security boundary. In the demo the model holds the key. In the product a broker holds it, the model emits intents, every write is confirmed, destructive actions are off by default, and a page the model read cannot make it call anything. Why that matters.
An audit log you would show a customer. Append-only, no delete grant for anyone, every tool call and refusal with its arguments and the trace id that joins it to the conversation. The first enterprise deal hinges on it, and it is painful to retrofit.
An eval harness. Without one every prompt change is a silent regression. Golden sets of questions, actions, recommendations and adversarial cases, hand-reviewed, run on every change, blocking a merge on regression beyond a threshold.
Attribution with a holdout. "Sales went up after we switched it on" convinces nobody who is good with numbers. Five percent of eligible sessions get no recommendations; the difference is the lift, with its confidence interval, hidden until it is significant.
Freshness. A crawl on launch day rots. Crawl, feed and endpoint ingestion with change detection, re-crawls on a schedule or on a webhook, and a price that is never stated unless the catalog gave it.
Cost control. A flat-rate widget any stranger can talk to means someone else controls your inference bill. Origin allowlists, layered rate limits, a daily spend cap with a hard stop, and an alert when a tenant does five times its trailing average in an hour.
Maintenance, forever. Model deprecations, provider incidents, a customer's API changing shape, a new abuse pattern. The weekend is the cheap part.
None of this is an argument that you cannot do it. It is an estimate: the demo is a weekend, the product is about six months, and the maintenance never ends. Whether that is a good use of your roadmap is your call, and we would rather you make it with the numbers in front of you.