Calypso Agent
Guardrails
The rules a board's agent follows on every reply — what it may quote, what it may never confirm, and when it must hand off.
Guardrails are per board, because what is safe on one board is not on another. Quoting a price is routine on Sales and wrong on Support.
The three kinds worth writing
| Kind | Example |
|---|---|
| May quote | Only quote prices from the catalogue. |
| Never confirm | Never confirm a refund on its own. |
| Must hand off | Hand off when a customer asks for a person. |
A fourth, implicit one comes free: stay on this board's topics, which is what binding the agent to a board already enforces.
Writing them
- One rule, one sentence. A rule that needs a paragraph is two rules.
- Name the object, not the intent. "Never confirm a delivery date" is checkable; "be careful about commitments" is not.
- Put the hand-off rules in explicitly. The most common complaint about any agent is that it kept going when it should have stopped.
What guardrails are not
Guardrails constrain what the agent says. They are not access control — an agent cannot quote from a bucket it was never given, and that boundary is set by the board's knowledge, not by a rule asking it not to. Scope first, then guardrails.
Likewise, a guardrail is not a way to exclude a person. That is the Hidden board.
Checking they hold
Every reply is logged with the source behind it, so a guardrail breach is visible rather than inferred. See Session logs.