Name and avatar
How the agent introduces itself and the photo the customer sees. This is the role it holds in your operation: salesperson, receptionist, broker, support analyst.
Hal-AI Agentic
The agent that actually handles the work — configured on a screen, not in a flowchart
A channel agent is the one talking to your customer. You define its name, personality, voice, and memory, then hand it real tools and a knowledge base of its own. From there, it decides what to do with every message.
This isn't a chatbot running a flow. It's an operational agent trained to work.
The same agent answers on WhatsApp, in your website chat, and on the phone.
Channel agents by the numbers
11
tools the agent can reach for inside a conversation
3
formats accepted in the Knowledge Vault: PDF, text, and image
1
same brain on WhatsApp, webchat, and voice
Setup
What you actually configure in an agent
Seven decisions on a screen. None of them is a conversation path — they're the traits of the worker who will do the job.
Personality and prompt
The text that describes who it is, what it may promise, the tone it uses, and where it stops. This is where your business rules live — written in plain language, not wired into decision nodes.
Voice
The voice the agent uses when it replies with audio and when it answers the phone. Same agent, same prompt, coming out as sound.
Linked channel
Which official WhatsApp number, which webchat, or which voice extension belongs to this agent. One agent can sit in more than one place without being rebuilt.
Administrator phone numbers
Which numbers the agent recognizes as the manager's. That's what separates someone asking about a price from someone asking for the weekly report.
Memory on or off
Long-term memory is a per-agent switch. On, and the agent remembers the customer next time. Off, and every conversation starts clean.
In the platform, those six decisions fit into a single tab. The screen below is the Personality tab of a channel agent's editor: the WhatsApp channel it is linked to, the name the customer sees, the internal description, the voice profile used in audio and on the phone, the temperature control between Precise and Creative, the long-term memory switch with the warning that it consumes more tokens, the phone numbers the agent recognizes as the manager's, and the system prompt — the text that describes the job and its limits.
The other tabs hold the rest of the agent: Memory shows what is retained per channel and lets you erase it, APIs lists your system's tools with method, authentication, and a usage counter, MCPs gathers the connected servers, and Web Chat Embed hands you the code snippet for the website chat. In none of them is there a flowchart, a decision node, or a scripted conversation path.
Agent Configuration
Create and modify the behavior of your AI agents.
Back to agents
Linked Channel
WhatsApp · Aurora Store Active
Agent Name
Aurora — Sales
Agent description
Handles the store's WhatsApp, checks inventory, and opens orders.
Voice Profile
Camila — clear voice, medium pace Play Sample
Temperature and Creativity (0.3)
Long-term memory
On — the agent remembers the customer between conversations
With memory on, the agent keeps a dossier on the customer (what they have asked for, preferences, history) and reads it at the start of every conversation. Off, it sees only the latest messages in the current conversation. Consumes more tokens.
Administrator Phone Numbers (WhatsApp Admins)
55 11 9xxxx-1201
55 11 9xxxx-4470
Add
System Prompt
You are Aurora, a salesperson at Aurora Store. Speak in American English, in short sentences.
Check inventory and delivery time with the tool before promising anything to the customer.
Never offer a discount that is not in the price table. If the customer asks for one, hand off to a person.
Tools enabled
api_get_inventory
api_post_order
GET /orders/status
api_del_order
The tools come from the APIs tab: each one has a method, authentication, a validation seal, and a usage counter.
Unsaved changes
Save Configuration
The row of enabled tools
Just below the prompt, the Tools enabled row is this agent's inventory of actions: checking inventory and opening an order are lit up, the status lookup is identified by its method, and the delete-order route is dark — deleting an order is not part of her job. At the bottom of the screen, the unsaved-changes badge waits for the manager to confirm before any of it reaches a customer.
Permissions change on one screen
Taking a permission away or handing it back means dimming a chip or lighting one up. When business policy changes — the agent may look an order up, but may not cancel it — the change happens on this screen, in minutes, instead of turning into an integration ticket in somebody's queue.
It is also the same screen for every channel this agent works on: one decision instead of three parallel settings someone has to keep identical.
The voice field sits next to the prompt
Voice Profile appears in the same tab as the prompt and the tools because the one taking the call is this agent, not a sibling of it: the same personality text, the same tools, and the same memory, live on the call.
In a flow-based automation, voice service is a separate product, with a separate script to write and keep current. Here it is one field on the form.
Why this matters
Configuring an agent takes minutes. Maintaining a flowchart takes all year.
In a flow-based chatbot, every new customer question becomes a new branch on the tree — and somebody has to draw that branch. Here, the new question meets the same agent, with the same knowledge and the same tools, and the agent handles it.
When a business rule changes, you rewrite a paragraph of the prompt or upload a document to the Vault. There is no map to redraw.
Flow-based chatbot
- Every anticipated question becomes a node
- Off script, it falls back to a menu
- Rule changes mean redrawing the map
- Forgets the customer with every conversation
Hal-AI channel agent
- Decides what to do at every turn
- Looks it up, searches, and acts when it doesn't know
- Rule changes mean editing the text
- Remembers the customer and what was left open
Tools
What the agent can do on its own
A tool isn't a menu button: it's an action the agent chooses to take when it makes sense for that conversation.
Run the conversation
Everything a human rep would do inside the chat, the agent does on its own initiative.
- Send media — photo, document, audio
- Send a contact card
- Send interactive options
- Request the customer's location
- Transfer to a human
Look up what it doesn't know
When the answer isn't in the prompt, it goes looking instead of making something up.
- Search the knowledge base
- Read a website
- Search the internet
And with your system's APIs registered, it can query the ERP, the patient record, or the order system mid-conversation.
Act beyond the present moment
Service doesn't end when the message does. The agent flags what has to come back around.
- Create and send a WhatsApp template
- Schedule a task
- Schedule a wait
The official template is what lets you reach a customer outside Meta's 24-hour window.
A tool in use
It doesn't say “let me check.” It checks.
When the agent needs a piece of data that only lives in your system, it calls the API you registered as a tool, reads the response, and keeps the conversation going with the number in hand. The customer never sees the seam.
The same holds in the other direction: the agent writes back to your system when the prompt allows it, and logs what it did in the CRM.
Knowledge Vault
Every agent's own knowledge base
The price of a consultation, the delivery window, the return policy, an internal procedure. What your company knows, the agent comes to know — without turning any of it into code.
You feed it in three formats
Upload a PDF, text, or an image, or write knowledge blocks directly on the screen. The content is indexed and becomes searchable for the agent during a conversation.
- PDF — procedure tables, manuals, template contracts, catalogs.
- Text — a block typed on screen, for the rule that changes every week.
- Image — a photographed menu, a flyer, a floor plan, a spec sheet.
Each agent has its own vault. The salesperson doesn't need to carry the finance manual.
And the agent feeds it too
The agent itself saves new knowledge mid-conversation and removes what turned out to be wrong. If a manager corrects a piece of information over WhatsApp, it goes into the vault and holds for the next customer.
That's the difference between a base that ages quietly in a repository and one the operation updates day to day, with no ticket and no release.
The knowledge base is what lets the agent answer about price, timing, and procedure without anyone programming a flow.
The screen below is the Knowledge Vault of the Aurora — Sales agent: search and format filters at the top, the uploaded files in the middle, and, at the bottom, the knowledge blocks typed straight onto the screen. Each file carries its own indexing state — the PDF catalog already indexed and available to search, the returns policy text still being processed, and the spec sheet image waiting in the reading queue. Only what is already indexed answers the customer; the rest joins the vault the moment it finishes, with no release and no ticket.
The two strips at the bottom show the vault in use: the lookup the agent ran against the base during a conversation today, and the block the agent itself saved after a manager's correction over WhatsApp.
Knowledge Vault
What this agent knows about the company. Each agent has a vault of its own.
+ Add knowledge
Meridiano line catalog
142 passages available to search
Returns and refunds policy
Follow indexing
Spec sheet — Meridiano armchair
Follow reading
Block · Delivery window
Capital and metro area: 3 to 5 business days after the order is confirmed.
Everywhere else: 7 to 12 business days. Made-to-order items always leave the factory on Tuesdays.
Block · Returns without a defect
The customer has 7 calendar days from delivery to request a return on an item with no defect.
The piece must be unused and still tagged. Pickup is scheduled with the carrier.
Block · Store hours
Aurora Store — Monday to Friday, 9am to 7pm; Saturday, 9am to 2pm. Closed Sunday.
Blocks typed straight onto the screen enter the index the moment you save. They are the right format for the rule that changes every week.
Agent lookup
search_knowledge_base("armchair delivery window") → Delivery window block · meridiano-catalog.pdf
Saved by the agent itself
In the index
save_knowledge_index("Meridiano line leaves the factory on Tuesdays")
What you're looking at
The size of the vault is in the counters
The numbers next to the filters say what this agent carries today: fourteen items, across PDF, text, and image. The catalog that is already indexed announces 142 passages available to search — a PDF does not sit in the vault as one whole attachment, it becomes passages the agent searches through.
On the right end of the bar, the agent selector swaps the entire vault: each agent sees only what belongs to it.
The advantage
The people who know the rule are the ones who maintain it
Upload the file or type the block and save: the rule that changed this morning is already answering customers this afternoon, with nobody translating policy into code first.
And because a lookup returns the passage rather than the document, the customer gets the sentence that answers their question instead of a manual attached for them to read alone.
Only on Hal-AI
The search is the agent's decision, not a step in a script
Look at the lookup strip at the bottom of the screen: it logs a call made from inside a live conversation, with three passages returned. Nobody marked “look it up in the base here” at some point along a path — the agent is the one that concluded, in the middle of the customer's sentence, that it needed to search before answering.
In a flow-based automation, the search only happens at the node where someone predicted it would; a question that arrives outside that point comes back as a menu.
- Looked up mid-turn, not before it
- Passage returned with its source in plain view
- The same tool serves text, webchat, and voice
Memory
It doesn't start from zero every conversation
Customer history is vectorized with embeddings, and a worker pulls stable facts out of it — episodic memory — along with customer traits, a read on mood, and a sense of time.
It's the same technology behind text service, voice service, and Squads. That's why a conversation still flows weeks later, and why the agent knows that this customer already complained about a delivery date back in March.
It's also the subject of the patent filed with the USPTO in 2025 — Autonomous Agent Memory Framework Inspired by Human Cognition.
What gets stored
Stable facts
What doesn't change message to message: the plan they signed up for, the delivery address, the restriction they declared.
Traits and mood
How that customer tends to talk, and what state they showed up in over the last few conversations.
A sense of time
How long ago something was said — the difference between a live loose end and a closed matter.
Manager control
The conversation memory screen shows what is stored per channel and lets you clear it.
Automatic schedule
The agent works even when nobody wrote in
Every agent has a schedule of its own: routines it runs at a set time, and waits it creates for itself inside a conversation.
Scheduled routine
A recurring task configured per agent, at a defined time. It wakes up, does what was agreed, and logs the result.
A wait created in the conversation
“Follow up with this customer in 2 days.” The wait comes out of the conversation itself, the moment the agent decides something needs a callback.
Every schedule in plain sight
The screen lists what's active per agent and lets you create, edit, and follow along. Nothing runs hidden from the manager.
Routines for the whole business belong to a Squad
A channel agent's schedule serves the conversation: the callback it promised, the reminder it agreed to, the check-in on that one customer. When the routine belongs to the business — sweep the ERP every morning at 7, decide who needs to be contacted, and send the channel agent to contact them — that work belongs to the orchestrator agent.
Copy and share
A good agent shouldn't have to be written twice
The agent that works becomes the starting point for the next one — inside your company or between companies on the platform.
Copy
Duplicate an agent you've already tuned and change only what differs: the branch, the language, the channel, a passage of its personality. What was right stays right.
Share
An agent can be shared with another company on the platform — and the other company has to accept. Nothing lands in anyone's account without the other side saying yes.
Useful for multi-unit groups, for franchises, and for the reseller program — the partner hands over a finished agent and the client accepts it.
Test chat
Talk to the agent before your customer does
The platform includes a simulator identical to WhatsApp so a manager can test the agent with no real number, nothing published, and no template spent.
- Same personality, same tools, same memory
- Test media, audio, interactive options, and transfers
- Test conversations stay separate from real ones in the CRM
- Adjust the prompt and test again on the spot
Once the agent behaves the way you want, link the channel. The same agent shows up on official WhatsApp, in your website webchat, and on the phone.
Service agent
online
Good morning — can I still exchange the item I received on Friday?9:12 AM
Good morning! I found your order. The exchange is still within the window. Would you rather drop it off at a store, or should I schedule a pickup at your address?9:12 AM
Pickup, please.9:13 AM
Pickup requested. I'll let you know as soon as the carrier confirms the date — and if I don't hear back by Thursday, I'll follow up with you.9:13 AM
Service agent
online
How many exchanges came in today?6:40 PM
I checked the system: today's exchanges already have pickups requested. One has been waiting on carrier confirmation since yesterday — want me to chase it?6:40 PM
Go ahead. And let the customer know.6:41 PM
Done. I scheduled another check for tomorrow morning.6:41 PM
The agent recognizes administrator phone numbers and answers the manager differently.
A human in the same chat
When a person steps in, the agent steps out — and comes back after
Transferring to a human is one of the agent's tools, but an operator can also take over on their own. On takeover, the platform pauses the bot in that chat automatically; when the operator closes out, the agent resumes. No parallel conversation, no lost history.
FAQ
Frequently asked questions
Do I have to build a flow or a decision tree?
No. The agent is given a personality, knowledge, and tools, and it chooses what to do at every turn of the conversation — including deciding that nothing needs to be done. There is no flowchart screen, no decision node, and no pre-mapped path for every possible thing a customer might say.
How does the agent learn about my company?
Through three paths that work side by side: the Knowledge Vault, where you upload PDFs, text, and images or write knowledge blocks right on the screen; your own system's APIs, registered as agent tools; and reading web pages during the conversation. The agent also saves new knowledge while it works.
Does the agent remember a customer from one conversation to the next?
If long-term memory is turned on for that agent, yes. History is vectorized, a worker extracts stable facts and customer traits, and the conversation picks up naturally weeks later. The memory screen shows what is stored per channel and lets you clear it.
Can I test the agent before connecting it to an official number?
Yes. The platform includes a test chat that simulates WhatsApp, with the same personality, the same tools, and the same memory as the published agent. Test conversations stay separate from real ones in the CRM.
Bring us a conversation that stalls today and we'll show you how the agent handles it
Tell us which conversation your operation repeats every single day. We'll design the agent that handles it with you — with the tools and the knowledge that conversation demands.