AI agents in XBASE
What an AI agent is in XBASE, the difference between Native and External, what you can configure, where it works and how it uses credits.
Updated on
An AI agent is an assistant that talks to your contacts on the company's channels, like WhatsApp. It replies on its own, based on what it knows about your business, and the team takes over the conversation whenever it wants.
In XBASE, agents are in AI Agents, in the menu. To create your first one, follow the step-by-step guide.
The two types of agent
| Type | What it is | For whom |
|---|---|---|
| Native | the XBASE AI reads your website, Instagram and your instructions and builds the agent for you | people who want a ready agent, without coding |
| External | an agent that already runs outside XBASE (for example, on n8n or your own API). XBASE sends each message to its address and delivers the reply to the contact | people who already have an agent and just want to connect it to the XBASE channels |
The type shows on each agent's card, next to the status. The External agent is configured outside XBASE: here you only keep the address, the token and the channels.
How the Native agent is created
When creating it, you enter the agent's website, Instagram and objectives, and you can write additional instructions. The XBASE AI:
- reads the website's home page and up to 7 internal pages;
- reads the public Instagram profile: name, bio and what shows without logging in;
- writes the agent's instructions, with a conversation script for each objective.
The screen shows the progress. At the end, it lists What the agent doesn't know yet and the sources that could not be read. Fill in these gaps in the agent's instructions.
The Native agent's AI model is set by XBASE: you don't choose it when creating the agent.
What you can configure
Open the agent in AI Agents. In the Native agent, the main areas are:
- Panel: the Connected Channels cards (where the agent works), Agent Instructions (the text that defines the agent's role, tone and rules), Intelligence Model, Observability & Analysis and Agent Objectives.
- Memory → Knowledge: files the agent consults to reply, like catalogs and policies.
- Learnings: improvement suggestions drawn from conversations, when automatic evaluation is on.
- Webchat: look and code of the website chat.
The instructions have a Version history: you can go back to a previous version.
In the External agent, the Panel shows the Agent webhook card (address, token and maximum response time, with the Test webhook button) and the channels.
Where the agent works
| Channel | Native | External |
|---|---|---|
| WhatsApp (Meta's official API) | not yet | yes |
| Website chat | yes | no |
| Instagram (direct messages) | not yet | no |
- Each number or account is handled by only one agent. One agent can handle several numbers.
- To choose the channels, use the Connected Channels card, in the agent's Panel: one switch per number. The Add channel button connects a new channel.
- Turning the agent on for a number that already has another agent asks for confirmation: the new agent starts handling the number, including ongoing conversations.
- Only people with permission to manage integrations can change an agent's channels.
- To chat with the agent yourself: in the Chatcenter, click New conversation, choose Agent, select the agent and click Create Conversation. The Title is optional: without it, the conversation takes the agent's name. The conversation is just between you and the agent, with no contact or channel, and its agent is already selected whenever you open it. Only Native and Active agents are listed.
To add the agent to your website, see how to add the chat to your website.
How the agent replies on WhatsApp
On each number, the How the agent replies block defines:
- Wait before replying (seconds): consecutive messages from the contact within this time are answered together. In the Native agent, it comes set to 10 seconds; in the External one, to 0 (replies right away). The maximum is 600.
- Format: A single message or Several short messages (one paragraph per message). In the second case, Meta charges each short message separately.
- Pause when a team member takes over: comes turned on. With it, the agent stops replying when someone on the team replies to the conversation.
The block is in the number's AI agent window, in Channels. Details on templates and the 24-hour window are in WhatsApp in Chatcenter.
When the agent stops and the team takes over
The agent does not reply to a conversation when:
- someone on the team takes over the conversation: in its header, the AI replies / Human selector shows who is handling it;
- the conversation has the "Atendimento Humano" (human support) tag;
- the agent is Paused. The status switch is at the top of the agent's page, and pausing asks for confirmation: the agent stops replying on every channel until you turn it on again;
- the workspace's AI credits ran out (see below).
To hand the conversation back, click AI replies in the conversation header. The agent starts replying again from the customer's next message. More details in WhatsApp in Chatcenter.
Agent status
| Status | What it means |
|---|---|
| Active | the agent replies on the channels linked to it |
| Paused | you turned the agent off; it does not reply on any channel |
| Draft | the Native agent has not finished being created, or the creation failed |
| Failed | there was an error with the agent |
Only an Active agent can be linked to a channel.
Credits and costs
- Native agent: each reply uses an AI model and uses AI credits from the workspace balance. Creating the agent with the XBASE AI also uses credits. Usage shows in Usage & consumption.
- External agent: processing happens outside XBASE and does not use AI credits.
- Out of credits: the Native agent stops replying. The conversation shows, only to the team, the "AI paused: out of credits" notice, and nothing is sent to the customer. When the balance is back, the agent replies from the next message on; messages left behind are not answered.
- WhatsApp messages: the agent's replies on WhatsApp also follow Meta's pricing, like the team's. See WhatsApp billing.
- Automatic evaluation: if you turn on Observability, each evaluation also uses credits. See Automatic conversation evaluation.
Plans and credit top-ups are covered in Plans, credits and billing and on the pricing page.
Rules and limits
- To create a Native agent, provide at least one source: the website, Instagram or additional instructions with 30 characters or more.
- Objectives: 1 to 5 per agent. The first one is the main one.
- Additional instructions: up to 4,000 characters.
- Instagram must be a public profile. A private profile, or one that only opens when logged in, cannot be read.
- External agent: the webhook address must start with https://, and the maximum response time ranges from 1 to 120 seconds (default: 60).
- Creating and editing agents requires permission to manage agents in the workspace.