Agent settings

This page is the reference for everything you can tune on an agent from the Dashboard: how it introduces itself, how it sounds or writes, what it can do and what it records from each conversation. Use it when you want to change something specific or know what a setting does before you touch it. Each agent only shows the sections for its kind: a voice agent has no Chat experience and a messaging agent has no Voice.

Before you start

• You need a role that can Create and tune agents, such as Agent design, Administration or Owner. With other roles the sections are read-only (“You have read-only access to this section.”) and Save doesn’t appear.
• Data it collects and Conversation reports also need permission to define the data that’s captured. If you can’t save them, ask whoever manages your team.
• Open the agent: in the menu, click Workers and then its name.

How a change is saved

Changes apply from the agent’s next reply or call. On a voice agent, they take a few moments to reach calls.

Each row says what state its section is in:

• Done: it’s set up.
• Pending: the agent needs it to serve people. There are only two: Identity & instructions on every agent, and Greeting on voice agents.
• Optional: it uses the defaults, which are fine to start with.

The sections at a glance

Agent foundation

What always applies, regardless of where it answers.

Identity & instructions

Change them whenever something permanent changes in how you work. How to write them: Write the instructions and Writing good prompts.

Language and time zone

Change the time zone only if the agent serves a site in another country or another time zone.

Call experience (voice agents)

How it sounds and what happens while it’s on the phone.

Voice

Search for a voice by name, listen to it with the play button and select it. The one it uses now shows first and selected; if it’s no longer in the catalog, it shows as Current voice. An agent with no voice chosen uses the default voice. The sample plays the original voice, without your settings: to hear the result make a test call.

Under Voice settings:

Change the speed if your patients ask it to repeat itself or calls feel slow. How a voice agent handles a call, from start to finish: Voice and audio.

Greeting

Welcome message: the first thing the agent says when it answers or places a call. It’s required, up to 2,000 characters. Example: “Hi, this is Sofía from Clínica Ejemplo. How can I help you?”.

Conversation pacing

Raise the Maximum duration if your calls, such as a confirmation with several steps, get cut off before they finish. Turn on Hide numbers entered on the keypad if you ask for sensitive details on the keypad. What happens when the person doesn’t speak: When the patient doesn’t answer.

Intelligence

Language model: the one that processes what the person says and decides what to answer and which tools to use. By default, Default (Arbolito Agile). See the options in The models.

Chat experience (messaging agents)

How it opens the conversation and how it behaves while writing.

Greeting

Under When someone writes you choose what a person receives when they open a conversation on WhatsApp:

With Send a fixed message you can turn on Always send the greeting, for notices everyone must receive. Examples and tips: Write the instructions.

Chat behavior

When to change them:

• Turn on the follow-up if many people leave conversations halfway, and choose a Close after inactivity longer than the follow-up: if the conversation closes first, the follow-up never goes out. See When the patient doesn’t answer.
• Turn on Pause AI when a person replies if your team also answers from the WhatsApp Business app. See Handle a WhatsApp conversation.
• Turn off Offer reply options only if your patients prefer to type freely.

Intelligence

The models

Both kinds of agent choose from the same models. All of them read images and PDFs.

If you have no reason to change it, keep the default. Try a more capable model when the agent has long instructions, many tools or makes mistakes in multi-step tasks, and test it before it serves patients: a more capable model can also take longer to answer.

Work & quality

What it can do, what it records, and how you review its work.

Skills (messaging agents)

A skill is a procedure the agent reads only when the conversation needs it, for example “check-availability”. It takes task-specific steps out of the main instructions. Skills are used on WhatsApp and in web chat.

1. Click Add skill.
2. Fill in the Name (short, lowercase, with hyphens), When to use it and the Instructions: the procedure, the checks and how to answer.
3. If you want, add References with Add reference: a Reference path, such as references/guide.md, and its Content. Mention that path in the skill’s instructions.
4. Leave Active on and click Save. You’ll see “Skills saved”.

An agent takes up to 32 skills, each with up to 16 references. A skill can only use tools the agent has already been assigned.

Tools and knowledge

This section holds:

• Escalation and support › Escalate to a person: in When should it escalate? you write the cases that need your team, with concrete criteria, and click Save escalation policy. An explicit request to speak to a person is always respected. On WhatsApp, after escalating, AI stops replying and only someone on the team can turn it back on. See Write the instructions.
• Escalation and support › Consult a supporting AI (messaging agents, on WhatsApp): with Enable support for blockers, before escalating the agent checks the reason and may consult a second AI once, which uses no tools and doesn’t take over the conversation. You can choose the Supporting model, write What should it review before escalating? and, optionally, Another Agent’s specialty. It has an extra cost per consultation, with its own limits: ask Arbol’s team to set it up with you. Off by default.
• Knowledge: the documents the agent searches to answer. See Give your agents knowledge.
• Integrations: your organization’s systems where the agent can look things up or make changes, and which of each one’s tools it uses. See The agent’s tools.

Data it collects

The information the AI extracts from each conversation, such as the preferred date or the insurer. What it captures shows under Captured data in each conversation and in the person’s Data.

1. Click Add field. You can start from one of the templates offered.
2. Write What to capture, for example “Preferred date”.
3. Click Show advanced settings and fill in:
   • Type of value: Text, Number, Date, Yes or no or List of options. With List of options, write at least two Allowed values, separated by commas.
   • Extraction instructions for the AI: exactly how to get the value from the conversation, for example “Extract the date the person prefers. Leave it empty if it wasn’t mentioned”.
   • Internal name: created automatically; change it with Edit internal name only if another system needs a specific name.
4. Click Save. You’ll see “Field saved”.

When you delete a field, it stops being captured in future conversations.

Conversation reports

What you want to know at the end of each real conversation, such as “Did the person schedule a follow-up?”.

1. Click Add report. You can start from a template, such as “Whether the reason for the contact was clear”.
2. Write What do you want to know?.
3. Under Show advanced settings, write the Detailed instructions for the AI, for example “Answer YES if the person confirmed a follow-up appointment, otherwise NO”.
4. Click Save. You’ll see “Report saved”.

The answers show under Evaluations in each conversation and in Reports. If the agent has at least one yes-or-no report, those reports decide whether each conversation is a Success or a Failure: see Understand and change the success rate.

What isn’t set on the agent

Settings Arbol’s team sets up

Some settings aren’t in the Dashboard. If you need them, ask Arbol’s team:

• A fixed message for when the person only says thanks or goodbye after a finished task, on WhatsApp.
• A monthly usage limit for the agent. If it reaches it, it stops replying on WhatsApp and conversations go to your team, with a notice for the person, until the next month or until the limit is raised.
• Answering a share of new WhatsApp conversations with another model, to compare it before switching.
• Requiring changes to the instructions, skills or model to pass a test against past conversations first (see Improve an agent’s replies).
• Having the agent search its knowledge automatically before every reply.
• Having the agent check again on its own a task your system left in progress, and answer the person when it finishes.

Common problems

For developers

All of these settings, and the ones Arbol’s team sets up, can also be read and changed through the API. See Agents.