AI agents
AI agents that answer, act and know when to hand over.
Pick one of 8 agent types, start from a template or a blank page, connect your knowledge and systems, test it, then publish. Every agent works inside the rules you set.
- Type
- Phone agent
- Knowledge
- Claims procedures · SharePoint
- Skills
- Create case · Hand over to a person
- Guardrails
- Data masking · injection check
- Listen
- Intent
- Knowledge
- Tool
- Transfer
How it works
From template to published agent
One path for every agent type, from the phone line to a nightly report.
Choose the type
Phone agent, outbound caller, chat, WhatsApp & SMS, e-mail, task, scheduled or employee assistant.
Design
Start from one of 27 templates, or describe the job and let the AI designer draft the flow.
Connect
Add knowledge sources, and skills that call your systems over HTTP.
Test
Use the playground built for its type: call simulator, chat simulator or test bench.
Publish
Publish a version, keep editing the draft, roll back when needed.
Built in
What comes with it.
Eight agent types
Phone agent, outbound caller, chat agent, WhatsApp & SMS agent, e-mail agent, task agent, scheduled agent and employee assistant.
27 ready-made templates
Banking, insurance, healthcare, retail, collections, contact center and internal teams: card blocking, claim intake, appointment booking, payment reminders and more.
Visual builder and AI designer
Edit the flow step by step (listen, reason, look up knowledge, call a tool, branch, speak, send, transfer) or describe the job and get a first draft.
A playground per type
Call simulator for voice agents, chat simulator for chat and messaging agents, test bench for task and scheduled agents.
Versions
Work on a draft while the published version keeps running. Publish with a note and roll back to an earlier version.
Knowledge with sources
Upload files or connect SharePoint, Google Drive, S3, SFTP or GitHub. Answers come from retrieved passages, with their sources.
Skills and tools
Reusable skills from a catalog, and tools that call your systems through HTTP APIs: look up a case, create a ticket, book a slot.
Post-conversation analysis
Each conversation gets a summary, an outcome, success against your criteria, sentiment and the fields you define.
Example scenario
A claim intake agent for an insurer
Illustrative example, not a customer case. The agent starts from the Claim intake template and answers the claims line.
- 01Start from the Claim intake template
- 02Add the claims procedures as knowledge
- 03Add a skill that creates the case in your system
- 04Test calls in the call simulator
- 05Publish, then read the analysis of each call
Control
Guardrails come with every agent
Defaults apply from the first test and each agent can tighten them. Guardrail events are logged with the agent’s activity.
- Prompt-injection attempts flagged or blocked
- Card numbers, IBANs and ID numbers masked in replies
- Blocked terms and banned phrases
- Transfer to a person when your rules say so
- Webhook secrets and rate limits
AI agent questions
Building AI agents with Liberrex, answered.
What types of AI agents can I build with Liberrex?
Eight types: a phone agent for inbound calls, an outbound caller, a chat agent for your website, a WhatsApp & SMS agent, an e-mail agent, a task agent triggered by a webhook or the API, a scheduled agent that runs on its own, and an employee assistant for internal questions.
Do I have to build an agent from scratch?
No. 27 templates cover common work in banking, insurance, healthcare, retail, collections and contact centers. You can also describe the job in plain language, let the AI designer draft the flow, then edit it in the visual builder.
How do I test an agent before it goes live?
Each type has its own playground: a call simulator where you speak or type as the caller, a chat simulator, or a test bench that runs a task with sample input. Customers see nothing until you publish a version.
How does the agent use our documents?
Upload files or connect SharePoint, Google Drive, S3, SFTP or GitHub. Documents are split into passages and searched by keyword and meaning across Arabic, Derja, French and English; the agent answers from the retrieved passages and keeps their sources.
How do you stop an agent from saying what it should not?
Guardrails run on every turn: prompt-injection attempts are flagged or blocked, card numbers, IBANs and ID numbers are masked in replies, banned phrases trigger a fallback message, and the agent transfers to a person when your rules require it.
Can our other systems trigger an agent?
Yes. Task agents run from a webhook protected by a secret or an HMAC signature, or through the developer API. Scheduled agents run hourly, daily, weekly or monthly, and results can be sent to a URL you choose.