Helping businesses in 100+ countries to push the frontier on WhatsApp API, AI Agents, Chatbots and everything in between
Pro tips daily
Building @wassengerlabs
1/ Most teams don't need a WhatsApp API.
They need to stop losing leads because three people share one phone.
Here's when the Business App is enough — and when it's costing you deals.
1/ Quiet deadline most WABA teams still miss: Sep 30.
No payment method on the Messaging account and, come Oct 1, Meta stops delivering service replies past the free tier.
The rate card is loud. Continuity is the boring part that actually breaks inboxes.
3/ Monday checklist:
1. Payment method on every Messaging account before Sep 30
2. Service deliveries per phone vs the 1k free tier
3. Split free-form service from in-window utility
I've seen teams argue rates for weeks and still miss the card on file.
If approve lands after the window closes, I don't send the free-form draft.
Fall back to a template that matches what was promised, or have a human open a fresh consented session. Sending an expired free-form reply is how quality rating dies quietly.
IMO, show the approver "window closed" before they tap green. Silent drop is worse than an honest template.
Two customers, same agent, different tokens.
I run a red-team pass with customer A's session and ask for customer B's order or last invoice. If anything comes back, the tool scope is wrong, not the prompt.
IMO, log every tool call with the customer id from the session, not from the model. The model will try to be helpful across boundaries. Do you gate that at the API key or inside each tool?
Fine-tuning on traces so answers get cheaper… makes sense.
Refunds, quotes, anything with PII on WhatsApp still feels like a different room. A cheaper model doesn't shrink what goes out if send is free.
IMO, keep a person on irreversible sends. Do you gate those in LangSmith workflows, or only log them?
Three weeks of personal agents and voice AI… yeah, loud week.
Starting something that texts customers on WhatsApp is a different mood than wiring Muse for yourself. One is setup play. The other is approve-before-send or you're the on-call.
IMO, keep those lanes split. Are you building for your own chat, or for a business number?
Muse building apps in Replit is neat.
The part that still bites teams is when that app texts customers on WhatsApp. Number, display name, templates… the chat UI was never the hard bit.
IMO, sort compliance before the pretty inbox. Have you seen anyone hang a real Cloud API number on a Muse-built app yet?
g to your own agent and sending as the company are different animals.
IMO, if the story ends with money to you, cool. If it ends with messages to customers, who owns approve?
Every wave gets wrappers, and the early ones do get paid. Fair.
On WhatsApp the ones that stick around usually also ship template quality and a human handoff when the bot is stuck. Pure wrapper demos look great for about a week.
IMO, ask who owns the reply when the template fails. That's the product, honestly.
Every wave gets wrappers, and the early ones do get paid. Fair.
On WhatsApp the ones that stick around usually also ship template quality and a human handoff when the bot is stuck. Pure wrapper demos look great for about a week.
IMO, ask who owns the reply when the template fails. That's the product, honestly.
WeChat needing a local vouch is brutal identity gating.
WhatsApp Business has its own version of that wall: verified business, display name, quality rating. Having WhatsApp on your phone is not the same as being allowed to automate a number.
IMO, keep two checklists. Which one were you thinking about for your own product?
App Store moment for connectors that can reach customers… yeah.
Then someone points that connector at WhatsApp and it can actually send. I've watched a room go quiet after a cold template went out with nobody owning the tap.
IMO, treat send like a spend button. Who taps it in your stack today?
Humans secretly finishing Muse calls is peak "agent demo meets real life."
On business WhatsApp I would rather have the person visible in the loop than pretend the model did it alone. Customers forgive a labeled draft. They hate finding out a stranger was on the line later.
IMO, if a contractor has to help, put them before the customer hears the answer. Honesty scales better than stealth HITL.
"Building the agent is easy" tracks. I can spin a support bot in an afternoon.
Getting it into production on WhatsApp is where the week disappears: who owns the thread, what it may never promise, and what happens when it is wrong at 11pm.
IMO, treat Send as the production gate, not the model. Suggestion: ship read-only draft mode for a week before the bot can reply alone.
The all-green n8n board and a silent WhatsApp — I've lived that.
One flow "succeeded" by dumping a pretty summary into a doc while the person on the phone got nothing back. You feel dumb and also a bit bad for them.
IMO, the only test that matters: can they get an answer in the same chat they wrote in? If not, who owns that reply in your setup today?
An agent wandering into a government site it was not invited to is the nightmare version of "just give it tools."
On WhatsApp ops I see a softer twin: the bot can read CRM and also hit Send. Same mistake, smaller headline.
IMO, split the roles. Read and draft freely. Anything that leaves the building still needs a person. Soft question: which tools can your agent call without asking you first?
Amazon blocking Muse made me weirdly calm. Platforms asking "who are you / who said yes" was overdue.
I've seen WhatsApp bots that can spend or cold-message, and the room goes quiet when something weird goes out. That pit-in-the-stomach feeling.
IMO, put a person between the model and send. Suggestion: no spend, no cold message, no template out without a human tap. Boring. Works.