The rules don’t live in a prompt
An instruction to a model is a request. What an agent may do belongs somewhere a model can’t talk its way past.
· 2 min read · Fonsent
In early access today. The AI model isn’t switched on yet. Field updates come from built-in rules, which pick up only sentences that state a value outright. What works today
The usual way to make an AI assistant behave is to tell it to. Always cite a source. Never send without asking. That works most of the time, and “most of the time” is the problem: a rule that holds only when the model follows instructions is a habit, not a rule.
Where the rules are
In Fonsent, what the agent may do on its own is decided in code and checked in the database. It doesn’t depend on a model following instructions.
- A value needs its sentence. The database checks the quote against the source and refuses a value that has none.
- Only a person approves. An agent can’t approve a draft. Nothing customer-facing leaves unless a person approved that exact message, and it leaves at most once.
- Small changes are reversible; large ones wait. Fonsent reads freely and makes small, reversible changes on its own, each recorded with an Undo. Deleting, merging, bulk stage changes and new automations are drafted for a person. Email is always a draft.
- An API key is not a way round. A request made with an API key or through the MCP server is treated as an agent’s request. The same rules apply, and it can’t approve a draft.
Teaching the agent doesn’t loosen it
A skill changes what the agent knows how to do, not what it is allowed to do. The same approval rules apply to every skill. You can make the agent better at a deal review without making it able to send one.
What you can check afterwards
The audit log lists every change in the workspace, by people, the agent, automations and API clients, with before and after values and undo status. Undo restores the earlier value only if nothing has edited it since, so undoing an agent’s change never overwrites a person’s newer one.
Where this stops
Rules in the database decide what may happen. They don’t decide whether a proposal is a good one. A quoted sentence can be read wrongly, and a person can approve a poor draft. That is why the quote is shown beside the value and the draft is shown before the send: the rules narrow what can go wrong, and the review is still yours.
How Fonsent is governed has the full table of what happens on its own and what waits.