What is this?
Capabilities are buttons the agent may press. Without them, it can only talk. With them, it can also update a contact, hand a chat to a person, or start one of your automations. Each one is a pre-built, tested action — turning one on doesn’t require writing any code, just deciding this agent should have that power. [SCREENSHOT: ai-agent-capabilities-tab — The Capabilities tab with a capability enabled and its confirmation/handoff-assignment options expanded]What can I do here?
Capabilities are grouped by area:
For any enabled capability, you can turn on Require customer confirmation — the agent asks “Shall I go ahead?” before it acts. For Hand off to Human specifically, set the Handoff Assignment: Unassigned, Fixed team member, Round robin, or Least busy.
How to use it
Turn on only what this agent actually needs
Each capability already comes with its own built-in description telling the model when it’s appropriate — you don’t write that part. You’re deciding which powers this particular agent should have, not how it should use them.Set Custom Field only writes to fields you’ve already created
The agent can save information a customer shares — a name, an email, a budget — into a custom field. It can only write into fields that already exist for this channel, and it checks the value against that field’s type (a real email for an Email field, an allowed option for a Dropdown). It cannot create a new custom field on its own — create the field first, then tell the agent when to save into it in Instructions, for example: “If the customer mentions their company name, save it to the matching custom field.”Trigger Automation
This is the capability people find least obvious, so here’s exactly how it works. The moment you enable Trigger Automation, the agent is handed a live, always-current list of this channel’s active automations — each with its own name and Description — every time a customer messages. This happens automatically. It has nothing to do with what you wrote in Instructions. The agent compares the customer’s message against each automation’s Description and starts whichever one fits. If nothing matches clearly, it’s told not to guess — it simply doesn’t call anything.What happens once a triggered automation is running
Control passes to the automation completely. If it includes an Ask step — collecting an email, a concern, anything — the agent steps entirely aside and isn’t consulted again until that automation reaches an end (Submitted or Not Submitted). Every reply the customer sends in that window goes to the automation, not the agent. This is deliberate: it’s what keeps a structured flow — say, collecting a name, email, and issue before raising a support ticket — from being interrupted halfway through. The trade-off is that the automation won’t recognize an unrelated question while it’s mid-flow; it treats whatever the customer sends next as the answer to its current question. Keep automations the agent triggers narrow and focused, and always connect the Not Submitted path so a customer who goes off-script isn’t left stuck — see Ask Questions.Triggering a sequence
Trigger Automation also works for Sequence-type automations — the agent starting one counts as an Assign Sequence action, including the same protection against enrolling a contact who’s already in that sequence twice.Troubleshooting / Technical Notes
- The agent isn’t using an automation I expected. Check that automation is Active, bound to this agent, and has a specific, clear Description — not the agent’s Instructions.
- It asked something unrelated to what I just said, mid-conversation. A triggered automation with an Ask step is running and has taken over — see “What happens once a triggered automation is running” above.
- No capabilities are listed, or the tab looks locked. Save the agent on the General tab first, then reopen it.