Actions: letting the agent look things up

Knowledge answers "how does this work?". Actions answer "where is my order?". An action is an HTTP request the agent can make to your systems during a conversation, so it can look up live data instead of guessing or handing off.

Setting one up

Open AI Agent → Actions and pick a template: Order status, Subscription status, or Custom HTTP. There is also a demo action that echoes back what it was given, useful for seeing the flow before wiring anything real.

Each action has:

• Name and a description of what it does and when to use it. The agent reads the description to decide when the action applies, so write it plainly: "Looks up the live status of an order by its number. Use it when the visitor asks where their order is."
• URL and method of the endpoint to call.
• Parameters the agent fills from the conversation, each with a short note on what it is.
• Headers for credentials. They are stored encrypted and never shown to the model.

Identity

None means the endpoint returns public data. Verified visitor means every request is signed, so your endpoint can trust which customer is asking. That is what lets the agent answer account-specific questions such as "what plan am I on?" safely. The signing recipe is on the widget settings page under Identity.

Write actions

An action marked Performs changes is one that alters something, like cancelling an order. These can be defined today and wait for a confirmation card in the conversation before they run, so nothing changes without the customer agreeing.

Testing and auditing

Every action has a Test button. You can preview as a specific customer by email or user ID to see exactly what the endpoint returns for them. Recent calls lists every execution, whether the agent made it during a conversation or you ran it from this page.