Instructions
Nine lines of plain language: the role, the tone, what it must ask before acting, and when to stop and pass the conversation on.
Product · AI Agents
An agent that would rather say “I don’t know” than invent.
Four layers decide every reply: what you told it to be, what it is allowed to read, what it is allowed to do, and what it must never do. You can read the full trace of any answer it gives.
- Answers with the source attached
- Egyptian, Iraqi, Gulf and Levantine
- Testable before publishing
Instructions
Nine lines of plain language: the role, the tone, what it must ask before acting, and when to stop and pass the conversation on.
Knowledge Base
PDFs, your website, and question-answer pairs. Each chunk is embedded and scored against every question, and the reply names its source.
Functions
Each function is an explicit permission with typed arguments: read an order, book a slot, escalate. Nothing is callable unless you enable it.
Guardrails
No answering outside the knowledge base, no unverified prices, no refund promises. Breaking a rule triggers a handover instead of a guess.
Four layers, in the order you configure them.
Nothing here is a black box. Every layer is a screen you can open, edit and test.
Knowledge Base
PDFs, your website, and question-answer pairs. Each chunk is embedded and scored against every question, and the reply names its source.
Functions
Each function is an explicit permission with typed arguments: read an order, book a slot, escalate. Nothing is callable unless you enable it.
Guardrails
No answering outside the knowledge base, no unverified prices, no refund promises. Breaking a rule triggers a handover instead of a guess.
Six things the agent does on its own.
Reads the dialect people actually type
Egyptian, Iraqi, Gulf and Levantine, mixed with English mid-sentence, plus Arabic typed in Latin letters. It replies in whatever it was written in.
Answers with a source attached
Every reply names the chunk it came from and its score, so a wrong answer becomes a content fix instead of a mystery.
Calls your systems with typed arguments
It picks the function and fills the arguments out of the sentence — order number, date, branch — then uses the response in its reply.
Remembers the contact, not the chat
Facts it learns are written to custom fields on the profile, so the next conversation on any channel starts informed.
Knows when to stop talking
Refunds, anger, legal questions, or anything with no source: it stops, assigns the conversation and passes the full history.
Testable before anyone sees it
Replay your real past conversations against a draft agent on a sandbox channel and read every trace before publishing.
What you configure, in plain terms.
What it isWhat it means for your setup
ModelChoose the model per agent. Faster models for high-volume support, stronger models for sales conversations that need nuance.
RetrievalDocuments are chunked and embedded. At query time the top chunks are scored; a minimum score is required before the agent is allowed to answer.
FunctionsDefined once with a name, a description the model reads, and typed arguments. Enable or disable per agent without touching the flow.
FallbackYou choose the sentence it says when it cannot answer, and whether that triggers a handover, a tag, or both.
TraceEvery conversation keeps intent, retrieved sources with scores, functions called with arguments and timing, and guardrail decisions.
Model names and limits change as providers update them. Check the current list in the app before you quote anything to a client.
What decides whether you trust it.
What stops it from making things up?
The grounding guardrail. Replies are assembled from retrieved chunks of your own content, and the trace shows which chunk and what score. If nothing scores high enough, it refuses and hands over.
How many agents can I run?
As many as your plan allows, each with its own instructions, knowledge base and functions. A sales agent and a support agent can share one inbox without sharing rules.
Can I see why it said something?
Yes. Open the conversation and read its trace — that is what you use to fix the instruction or the document, rather than guessing.
Build the agent, then try to break it.
Start on the free plan, point it at one document, and ask it the questions your customers actually ask. Publish when it stops surprising you.