Agents
Five assistants, and not one of them publishes
An agent that publishes dental marketing on its own is not an assistant. It is a way to generate a violation in the practice’s name, because the person a regulator penalises is the registered dentist and not the software vendor. Here agents produce drafts. Publishing stays an act a person performs — and there is no automatic mode, not even one shipped switched off.
Writing depends on a model, and the model depends on a server we do not have yet. The screens, the gate and the review already work. What is missing is named below.
The five, and what each one does
Not five chat boxes with different names. Each produces one specific thing, and the fifth exists to make the other four usable in a practice at all.
Partner
Answers questions about running the practice and suggests what to do next. Produces a conversation and a summary worth keeping. Writes nothing a patient reads.
Blog
Writes articles about the treatments the practice performs, for people searching. Produces a draft with a title and body — always for review.
Outreach
Prepares the reply to someone who asked for contact through the site and has not been answered. One draft per patient, and it only leaves after you read it.
Reviews
Follows what is said about the practice publicly and prepares the response. One draft per review.
Compliance
Reads everything the practice publishes against the dental advertising code and returns a list of what has to change, with the reason for each item. It is what makes the other four usable.
Three rules, and none of them is a setting
These are not safe defaults someone can change later on a preferences screen. They are the shape of the contract: the field that would allow the opposite does not exist.
Every draft is born marked as agent-written
Not an optional field someone forgets to fill, and not a boolean left false. It is a single-valued literal in the type: there is no path to constructing a draft without the mark. And it renders ABOVE the text, because a reader needs to know what they are reading before they read it.
- The mark travels with the draft, not with the screen that shows it.
- Machine text is never attributed by default to a person who did not write it.
- Who approved it, and what they changed, is recorded alongside.
The same ruler as text you write yourself — literally the same one
An agent’s text runs through the same function that checks the copy a practice writes by hand. Not a copy of the rule adapted for agents: the same code, in the same place, so the two cannot drift apart over time. An agent exempt from the ruler would be the fastest way to publish a violation that looks like process.
The two ways to put an agent in a practice
The difference does not show on the day the agent is right. It shows on the day it is wrong — and that day decides whose fine it is.
| — | Publish on its own | Draft, then a person |
|---|---|---|
| Who signs the text | The practice, without having read it. | The practice, having read it — and the record keeps who did. |
| When the ruler rejects it | It is already published. The fix comes after the public saw it. | It never reaches approval: the button stays disabled until the item is fixed. |
| If you want to switch on automatic | It is a toggle in settings. | There is no toggle. The field does not exist in the contract. |
| How much work it saves | All of it. Including the reading. | The blank page, which is the expensive part. Reading two paragraphs is not. |
What does not happen yet, and where the ruler applies
Two limits, both stated rather than discovered later. First: the agents do not write yet. The server the model runs on depends on an infrastructure quota we have requested and not received; what already works is the half that decides whether this is safe in a practice — the screens, the authorship mark, the advertising gate, the review queue and the record of the decision. Second: the ruler itself was written for Brazilian dental advertising and its term list is in Portuguese. In a market whose rules have not been verified against a primary source or local counsel, the gate is closed rather than permissive, and nothing patient-facing publishes there — the same rule the compliance gate follows everywhere else on this site.
About the agents
Can automatic mode be switched on later?
No. It is not an option that ships disabled: it is the absence of the field. Enabling it would mean changing the contract between parts of the product, which is an architecture decision rather than an account preference.
Can the agent publish if I authorise it in writing?
Not through the product. Responsibility for dental advertising sits with the registered clinical lead, and a prior authorisation does not move that onto software. If that ever changes, it changes with legal advice and with this line rewritten.
Who is the author of the article on my site?
The practice publishes under its clinical lead’s responsibility, and the draft carries the record that an agent wrote it and a person reviewed it. We do not attribute machine text to the name of someone who did not write it.
What if I want to rewrite all of it?
That is the common case and the product assumes it: the review field asks what you changed or why you refused. Refusing is a normal, recorded outcome, not a failure of the flow.
The right question is what happens the day it is wrong
In the demo we open a draft the ruler rejected on purpose and show what the screen does.