If you use WhatsApp on Android, you might already have quiet access to WhatsApp AI agents — a beta feature that lets you drop up to five outside AI bots into their own private chat threads. You set one up with a name, a photo, and an API key, then talk to it just like any other contact. There’s a catch: unlike your regular chats, these ones aren’t end-to-end encrypted.
- Up to 5 third-party AI agents per WhatsApp account, each in its own dedicated chat
- Spotted in WhatsApp beta for Android build 2.26.35.3, rolling out to a small slice of testers
- Setup needs an agent name, a profile photo, and a WhatsApp-generated API key — no coding
- Agent chats are not end-to-end encrypted; regular WhatsApp chats are unaffected
- First reported by WABetaInfo on September 4, with no public rollout date from Meta yet
How Setting Up a WhatsApp AI Agent Actually Works
The flow is simpler than you’d expect from anything involving an API key. Inside a new settings menu, you give your agent a display name and pick a profile photo, the same way you’d set up any contact card. WhatsApp then hands you a unique key tied to that specific connection.
You copy that key over to whatever platform is actually running the agent — a developer’s own server, a no-code bot builder, whatever they’ve built. Once it’s pasted in, the agent shows up as a chat thread inside WhatsApp, and you talk to it exactly like you would a person. Ask it something, it answers, all inside the app you already have open all day.
Right now this only works one agent, one chat, one direction at a time. There’s no group chat support, and an agent can’t see anything outside the single thread it lives in — not your contacts, not your photos, not your other conversations. Remove the agent and the chat history stays put, but the live connection to the outside service is cut immediately.
Why These Chats Skip End-to-End Encryption
This is the part worth slowing down on. Every normal WhatsApp conversation you have is end-to-end encrypted, meaning not even Meta can read it. Agent chats work differently on purpose.
For a bot to answer you, whatever’s running it has to actually read your message — there’s no way around that. According to WABetaInfo, messages sent to an agent travel through Meta-managed infrastructure on behalf of the developer, which is what makes the reply possible in the first place. It’s the same basic trade-off you already accept when you message Meta AI directly inside WhatsApp; this just opens that door to outside developers too.
None of this touches your other chats. A friend’s messages, a group thread, a call — all of that keeps its normal encryption. The exception is scoped tightly to the one agent thread you deliberately created.
What Works Today — and What Still Doesn’t
Because this is an early beta, the feature list is short and the gaps are obvious once you look for them.
- Works: text-based back-and-forth with an agent, custom name and photo, up to five agents connected at once
- Works: regenerating the API key if you want to swap which service powers an agent
- Doesn’t work yet: group chats, voice or video with an agent, or sharing media inside the agent thread
- Doesn’t work yet: any official agent from OpenAI, Google, or Anthropic — none have been announced as WhatsApp partners
That last point matters more than it looks. This is plumbing Meta is laying down, not a storefront of ready-made bots. Anyone building an agent today would need their own hosting and their own AI model behind the scenes — WhatsApp just gives it a place to live.
Who’s Likely to Build the First Agents
Reports describe OpenAI, Google, and Anthropic only as the kind of model providers a developer might use to power an agent — not as confirmed partners with a chat ready to go. Meta hasn’t named a single launch partner.
That said, Meta clearly isn’t waiting around for outside developers. It already ships its own agent, Meta Muse, and Google has been pushing Gemini onto every surface it can reach, including a dedicated Gemini app for Windows. Sticking a general-purpose agent hub directly inside the world’s most-used messaging app fits that same pattern: whoever gets a working agent into your WhatsApp chats first gets a huge head start on everyone else building the same thing.
| Chat Type | Encrypted? | Sees Your Contacts? | Group Support |
|---|---|---|---|
| Regular WhatsApp chat | Yes, end-to-end | No | Yes |
| Meta AI chat (built-in) | No | No | No |
| Third-party WhatsApp AI agent | No | No | No, not yet |
Frequently Asked Questions
Is the WhatsApp AI agents feature available to me right now?
Almost certainly not yet. It’s limited to a small number of Android beta testers on WhatsApp build 2.26.35.3, and Meta hasn’t given a date for wider release.
Do I need to know how to code to add an agent?
To use one someone else built, no — you just enter the API key they give you. To build your own agent, yes, you’d need to host it somewhere and connect it yourself.
Can an agent read my other WhatsApp chats or contacts?
No. Each agent is confined to the single chat thread you create for it and can’t see your contact list, other conversations, or shared media elsewhere in the app.
Will this come to iPhone too?
Every report so far points only to Android beta testing. Meta typically brings features to iOS afterward, but nothing’s been confirmed either way.
Is it safe to use if it isn’t end-to-end encrypted?
It’s no less safe than talking to Meta AI already is inside WhatsApp, but treat it like any other conversation with a company’s server behind it — skip sending anything you wouldn’t put in a regular email.
Honestly, this feels less like a finished feature and more like Meta staking out territory before OpenAI, Google, or anyone else builds their own front door into your group chats. If you’re an Android beta tester and it shows up for you, it’s worth poking at — just don’t treat an agent thread with the same trust you’d give a message from your actual contacts.
Leave a Reply