Agent roles and agent rules: what the agent may do, and the limits it cannot cross
The agent role sets the persona, what the agent may do and its starting goal. Agent rules block actions until details are captured, or add instructions.
One agent role is active: its System prompt is the persona, What it may do is a hard allowlist, and Goal it pursues is where conversations start. Agent rules add limits on top: Block tool until fields present and Require human before tool are checked in code before every action; Append instruction is guidance the AI usually follows. Topic hand-overs live on the Human escalation tab.
The agent role
| Field | What it does |
|---|---|
| System prompt | The persona: who the agent is, how it sounds, what it should and should not talk about. Guidance. |
| What it may do | The capabilities the agent may use — read.* for every read, write.task for tasks only. Anything not listed is refused in code, whatever the AI decides. |
| Goal it pursues | The goal a new conversation starts on. Make it your opener (a stance-only goal), not the goal that collects details. |
Only one role is active at a time. Switching roles changes persona, permissions and starting goal together.
The four rule effects
| Effect | What it does | Example |
|---|---|---|
| Block tool until fields present | The agent may not use the chosen capability until the listed fields have values. The agent is told which field is missing, so it asks for it. | No meeting is booked before a work email is captured. |
| Require human before tool | If the agent tries the chosen capability, the action is refused and the conversation is handed to a person. | Any change to a deal's value goes to a person. |
| Append instruction | A sentence added to the agent's instructions, optionally only when a condition on the customer's fields holds. Guidance, not enforced. | Always mention free cancellation up to 48 hours before. |
| Force a person on topic | Created on the Human escalation tab: a topic that always goes to a person. | Refund disputes and payment complaints. |
How the checks run
- Before every action the agent proposes, the gate checks the role's allowlist, the goal's Allowed tools and every active rule. A refused action is reported back to the agent as refused, never silently dropped.
- Rules match capabilities, not tool names: a rule on
write.opportunitycovers every field of a deal;write.opportunity.update.valuecovers only its value. - When a rule takes effect, the thread's automatic events show a Rule line naming it and what it did.
- A change to the agent's own rules or settings that the agent proposes is never applied directly — it lands in AI Suggestions for a person to accept.
My Append instruction is ignored sometimes.
It is guidance. If it must hold, make it a gated rule — block the capability until a field is present, or require a person before it.
Which rule should stop the agent quoting an unlisted price?
None is needed: the agent may only state a price that is in the catalogue, checked in code on every reply. Add the product to the Catalogue with its price.
Related
Stuck? Write to help@linkubit.com with the screen you are on and what it says. Not using Linkubit yet? Start free — fourteen days of Pro, no card.