Agent settings

Two kinds of setting live here, and they save differently. Persona, guardrails and the escalation alert stage as a draft and change nothing until you publish. Handoff rules and tools apply the moment you move them.

Each card says which it is.

Persona

The agent's name, its tone — friendly, professional or concise — and the languages it replies in.

Leave languages empty and it answers in whatever language the customer wrote in, which is usually what you want if you sell across borders.

Custom instructions are free text for the things tone does not cover: how to sign off, what to call your product, whether to use the customer's first name.

Guardrails

Topics the agent must refuse, and instructions it always follows. Written in plain language.

Use them for the things that are wrong to answer rather than merely unknown — medical or legal advice, competitor comparisons, anything about pricing you would rather a person handled.

Handoff rules

The moments the agent stops trying and fetches a person. Each is a switch:

• The customer asks for a person. Leave this on. An agent that argues with someone asking for a human is the single fastest way to lose them.
• The customer sounds frustrated. Detected from the conversation.
• The answer is not in your sources. The honest one — better a person than a plausible invention.

You can also have the agent tell the customer it is fetching someone, rather than going quiet mid-conversation.

Follow-ups

When a conversation is waiting on the customer, the agent can check back once — in their own language — instead of letting it go stale. Set how long it waits and what it says.

Tools

Tool servers connected over MCP, and which of their tools the agent may use. Turn on the ones it should have and leave the rest off.

Gate anything with consequences. A gated action is proposed rather than executed: the agent shows what it wants to run and a person approves or denies it.

Publishing

While a draft exists, a banner offers Publish changes or Discard. Customers keep seeing the live version until you publish, so nothing changes underneath a running conversation.

Pausing

Pause AI stops the agent answering on every channel — new conversations wait for a person. To stop it on one thread only, use the per-conversation pause in the inbox.