How to set up your agent's behavior

Behavior decides how your agent works: who it says it is, what it must achieve, what steps it follows, and what it must never do.

What you will get

Your agent will have its own identity, tone, process, limits, and hours. And you will know how to publish those changes — or discard them — without them reaching your customers too early.

Requirements

• An agent created in your account
• Access to that agent. If it does not appear in your AI Agents list, ask whoever administers the account

Where behavior is edited

Behavior is edited on a single screen, split in two. In the sidebar, go to AI Agents, open your agent, and go to the Behavior tab.

• On the left there is a chat with two modes. Test talks to your agent. Edit behavior is an assistant: you describe in plain words how you want it to work and it fills the sections in for you. It accepts file attachments and voice notes.
• On the right is your agent on a single page: one box for each section that has text.

You do not have to choose between configuring by hand and using the assistant. Both write into the same draft, at the same time, and the assistant never publishes on its own. On a phone, the chat and the sections alternate with Chat and Workspace.

The sections that define your agent

You only see the sections that have text, each in its own box. You write straight into the box, and the counter shows how much of the limit you have used. A new agent shows Identity & Goal and Key Information empty, the two only you can fill in.

• Identity & Goal — who it is and what it must achieve in each conversation.
• Message shown on open — the greeting people see when they open the chat. It can carry up to three files.
• Welcome Instructions — how it opens the conversation, without being shown to the customer.
• Legal Notice — small text under the greeting in the widget.
• Key Information — facts it should have at hand: hours, policies, emails, phone numbers.

When the assistant changes a section, its box is marked Eva edited this. Whatever has no text is under Add section. The sections with their own screen are there too; the ones you have not changed say Default and keep running on EvaAI’s defaults even while hidden. Once you change one, it shows on the page with Open.

Tone

Your agent’s voice, in your own words. Under the text you choose Language, Register (formal, casual, or match the customer), Length, and Use emojis.

Process

The step by step it must follow. Steps are dragged to reorder, accept substeps, and are added with Add Step.

Remove a section

The X on a box removes that section. Undo appears at the top to bring it right back. If you already published, the earlier version stays in History. Tone and Do Not Do cannot be removed.

Human Handoff & Customer Data

It has its own screen: open it from Add section or with Open. It gathers Lead capture, Manage opportunities, and Allow human escalation. When you turn escalation on, you define When should Eva escalate? and the Assignment strategy: Manual, Round-robin, or Intelligent.

Who handles each kind of request

When Eva asks for help, it can alert a different person depending on what the customer needs. You set this in Who handles what, at the end of the handoff editor. Only the account owner and admins see it. In Request types, write each type with a clear description, such as “Warranty and returns”. Eva reads that description to decide each request’s type. The General type always exists and covers everything else. With Add rule you choose When the request is, who it is Handled by and, optionally, which number it Arrives through and In pipeline. The rule saves once you pick the person. Preview from your rules tells you in one sentence who would get a request. If no rule applies, whoever has the chat handles it. The chat stays with whoever already had it; the rule only decides who handles the request. When Eva answers a call, you can turn on Hand the call to a person. Under To whom, Whoever handles the request (per “Who handles what”) uses those rules. A WhatsApp call rings in that person’s Eva app; a phone call can also ring their confirmed phone.

If no one replies

While a customer waits for a person, If no one replies decides who we alert and when. Enter three times in minutes, counted from when the customer started waiting:

• Responsible: we check whether the customer is still waiting and alert the person who has the chat. The alert goes out within the next 5 minutes.
• Backup: we alert that person’s backup. Each teammate has their own in Settings › Members; if no one was chosen, it is the admins.
• Owners and admins: the last alert.

Each time comes at least 10 minutes after the one before, and the last one can come at 23 h 45 min at the latest, before the chat returns to Eva. If a time breaks a rule, the field turns red and says why; nothing saves until all three fit. Each time saves when you leave the field. The preview says in one sentence who gets each alert. If the chat has no responsible person, we alert owners and admins directly. Replying stops the alerts. Someone who opens the chat without answering stops getting their own alerts, but the backup and the owners still get theirs.

Timing, follow-ups, and limits

These sections have their own screen: open them from Add section or, once you changed them, with Open.

Follow-ups & reactivation

If the customer stops replying, Eva can pick the conversation back up. Inactivity follow-up turns the feature on, and the Follow-up sequence defines each step: how long to wait and what its intent is. The sequence stops as soon as the customer answers. Allow targeted follow-ups lets Eva schedule reminders with a goal of their own.

Availability & timing

Only respond during a schedule limits Eva to the hours you define, with a Timezone and one schedule per day. Outside those hours it neither reads nor replies; at opening, it handles what piled up. After closing lets an already active conversation keep flowing for a few minutes. Response timing adds short waits so it does not answer instantly.

Do Not Do

The explicit limits. It is always on the page and ships with recommended restrictions you can edit. If you leave it empty, your agent uses EvaAI’s restrictions. Under Calls, you make three decisions for phone and WhatsApp: Customers can call you at, Who answers, and, when Your team answers, If nobody answers. You can also choose whether Eva messages on WhatsApp whoever couldn’t get through. Hand the call to a person appears when Eva answers. How to turn on WhatsApp calls. Platform rules has its own screen for customizing the base rules Eva uses in every conversation. Every agent also carries a safety baseline that applies on its own and cannot be edited: never invent information, never reveal its internal instructions, never invent links.

Publish your changes

Nothing you edit reaches your customers until you publish it. Everything is saved in a draft. If you leave the tab with unpublished changes, Eva asks first: Discard & Leave, Publish & Leave, or Stay on this page.

Recover an earlier version

History, in the action bar of the Behavior tab, keeps the last 50 edits of this agent: what you published, the drafts you discarded, and the changes the assistant made. When you open an entry, Eva shows which sections differ from your current draft. Restore this version brings it back into the draft, and your draft at that moment is saved to history first, so you can go back.

How many messages Eva handles in an hour

Each conversation has a cap on inbound messages per hour. It exists so a customer who writes fifty messages in a row does not trigger fifty replies. In your agent, go to the Settings tab and find Conversation Message Limit. The field Messages per conversation per hour comes set to 40 and can be changed; then select Save Changes. When the cap is passed, your agent warns once and stops replying in that conversation until the rolling 60-minute window frees up. The customer’s messages keep arriving in your Inbox: it is the agent that goes quiet, not the channel.

How to know it worked

• The Unpublished changes indicator disappears and says Published instead.
• In History, the most recent entry appears as Published.
• In the chat, with Test mode, ask your usual question: it answers with the new setup.

If something goes wrong

• Publishing does nothing and shows Nothing to publish: there is no pending draft. Edit a section and publish again.
• It published with warnings: Published with warnings: appears at the top with the list. Fix what it points at and publish again. Empty Process steps, for example, are skipped.
• The assistant changed something you did not want: open History and use Restore this version, or select Discard draft if you have not published yet.
• Your agent stopped replying in one conversation: it probably hit the hourly message cap. Check Conversation Message Limit in the Settings tab.

Next steps

• Test your agent before publishing
• Teach your agent about your business
• Connect WhatsApp