AI Agents
An AI agent answers customers for your team. You write how it should behave in a playbook, give it knowledge to answer from, choose the functions it may run (such as booking an appointment or creating a ticket), and deploy it to Inbox channels, Tickets, phone lines, or Chat.
Agents live in Settings → AI → AI Agents. AI work is paid from your AI balance.
Fox and your agents
Section titled “Fox and your agents”Fox is Workstation’s built-in AI assistant. It appears in the agent list with a System badge. You don’t build Fox; its page only controls which tools it may use, which actions need approval, and shows its activity. Fox is always available to teammates in Chat and acts with the permissions of the teammate who asks.
Every other agent is one you build.
Create an agent
Section titled “Create an agent”- Open Settings → AI → AI Agents and select New agent.
- Choose where it works: Chat, Calling, or Chat & calling.
- Start from scratch or a template, such as Receptionist, Customer Support, Lead Qualification, Appointment Booking, Shopping & Orders, Order Status, Lead Follow-up, or Appointment Reminder.
- Give it a name and pick Content Hub articles it can answer from.
- Optionally connect it to inbox channels now, then select Create agent.
Quick setup asks only for the essentials. Advanced setup adds identity, tone, language, and fallback behavior.
You can also type a request in the Ask Fox to build or change an agent bar and let Fox draft it.
The agent editor
Section titled “The agent editor”Opening an agent shows its editor. The playbook is on the left, and settings are grouped on the right.
Playbook
Section titled “Playbook”The playbook is one document that tells the agent how to behave: who it is, its tone, working hours, what to ask for, guardrails, and when to hand off. Use Add section for ready-made sections in English or Arabic, such as Human handoff, Book an appointment, or Lead capture.
The toolbar inserts building blocks into the text:
- Insert variable — a value such as the customer’s name
- Hand off to a teammate
- Apply a tag
- Collect customer details
- Insert function — run one of the agent’s functions
- Send a saved reply — a fixed macro, or Best match
- Insert flow — a multi-step procedure the agent follows exactly
Keep facts such as prices and policies in Knowledge, not the playbook. Long playbooks dilute the agent’s attention; the editor warns you as you approach the limit.
Generate from conversations writes or improves the playbook from how your team actually handles conversations. Choose Quick (a few dozen conversations, about 30 seconds) or Deep (up to a few hundred, runs in the background), a period, and channels. The first run drafts a playbook; later runs suggest additions, replacements, and contradictions between the playbook and what the team really does. Nothing is saved until you review and save it. The estimated cost is shown before you start.
Knowledge
Section titled “Knowledge”Knowledge is the set of facts the agent answers from. Select Add content and choose:
- Article — from Content Hub
- Uploaded file — PDF, DOCX, TXT, MD, CSV, or JSON, up to 20 at a time
- Website sync — crawl your website, with page limits and pages to include or exclude
Each source shows its status (for example Syncing, Indexing, Ready, or Failed). Website sources can resync automatically daily, weekly, or monthly. Knowledge changes take effect immediately.
Model, functions, and behavior
Section titled “Model, functions, and behavior”| Section | What you set |
|---|---|
| Model / Models & voice | Chat model, calling model, voice, and language |
| Functions | Actions the agent can run, such as creating or updating a contact, booking an appointment, creating an order or ticket, sending a WhatsApp template, or managing tags. Calling agents also get call functions such as End Call and Call Transfer. |
| Chat Settings | Let the agent end a chat, and an inactivity timeout |
| Speech Settings | Calling only: interruption sensitivity, response wait time, background sound, and pronunciation |
| Post-Chat Extraction / Post-Call Extraction | Details to pull out after each conversation, such as a summary or sentiment |
| Behavior & guardrails | Default phone country, external tool connectors, who owns records the agent creates, voice-message transcription, image analysis, and WhatsApp template fallback |
| Handoff to a person | Extra handoff rules, the message the customer gets, who receives the conversation, whether the agent keeps helping until a teammate replies, and a case summary for the teammate |
| Deployment | Where the agent works. See below. |
The Workflow view shows the order of steps around each conversation: functions that run before, knowledge, extraction, and functions that run after.
Test and publish
Section titled “Test and publish”Select Test to chat with the saved draft, or talk to it with your microphone. Test in inbox starts a test conversation in Inbox, where you write as the customer and choose Simulate actions or Run real actions.
Until you publish a version, every save goes live (Auto-publish). After the first Publish, edits wait as Unpublished changes while customers keep using the published version. The versions panel lets you view old versions and create a draft from one. Knowledge and deployments always apply immediately.
The header switch sets the agent Live or Paused.
Deploy an agent
Section titled “Deploy an agent”In Deployment, select Add a deployment and choose a target:
- Inbox channel — chat, email, WhatsApp, and social channels
- Tickets — helpdesk tickets
For each deployment, choose What it does:
| Mode | Behavior |
|---|---|
| Suggest a reply | Teammates see a suggested reply to use or ignore |
| Auto-reply | The agent answers customers itself (Inbox only) |
| Internal note | The agent adds a note for the team |
| Action only | The agent runs functions without replying |
| Triage only | The agent classifies and routes |
You can limit a deployment to selected channels and choose whether it handles direct messages, public Facebook and Instagram comments, or replies privately to comments.
Calling agents are not deployed here. Add them to a phone number’s call flow with the AI agent step. See Calls.
Chat coworker
Section titled “Chat coworker”The Repzo Chat card deploys the agent as a coworker in team Chat. Set its display name, whether teammates can message it directly, and which channels can @mention it. It acts with the permissions of the teammate who asks, and teammates approve its actions from cards in Chat.
Working alongside the agent
Section titled “Working alongside the agent”In Inbox, teammates can Take over a conversation from the agent, Resume AI, send it a private Instruct AI note, and rate its replies. Conversations it hands over appear under AI handoffs with a brief explaining why.
Improve an agent
Section titled “Improve an agent”The editor’s ⋯ menu opens three views:
- Feedback — replies teammates rated Good reply or Bad reply, grouped by what went wrong. Improve agent on a group proposes fixes. You can also turn rated replies into test sets, and review questions the agent couldn’t resolve.
- Analyze — run counts, answer and handoff rates, average cost, saved test cases, and evaluation datasets you can run and compare across versions.
- Activity — recent runs with the reasoning behind each answer, tool calls, handoffs, a queue of actions waiting for Approve or Reject, and cost traces.
From a single conversation, More actions → Improve the AI agent in Inbox analyzes what went wrong and suggests changes to knowledge, guidance, flows, or the playbook. Select Apply and then publish when ready.
Permissions
Section titled “Permissions”| Permission | Allows |
|---|---|
ai_agents.view |
See agents and their activity |
ai_agents.create, ai_agents.update, ai_agents.delete |
Build and change agents |
ai_agents.manage_knowledge |
Add and remove knowledge sources |
ai_agents.deploy |
Deploy agents |
ai_agents.approve |
Approve actions the agent asks to run |
Owners and admins have all of these by default; managers can view.