# We gave our AI agents a real Android phone instead of APIs. Here's what broke first. — retained record

> Retained reddit record from reddit:AI_Agents.

Canonical: https://highsignal.app/data/records/8bc62aff68e0cf22

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:AI_Agents)  published 2026-10-31 06:23 ingested 2026-11-01 02:36

[original source ↗](https://www.reddit.com/r/AI_Agents/comments/1xk2uqp/we_gave_our_ai_agents_a_real_android_phone/)

## retained text



```
For the last few weeks I've been running a setup that sounds dumb on paper: a rooted Pixel with its own SIM, sitting on a desk in Milan, driven by AI agents from several different projects. They send WhatsApp and SMS, manage groups, browse the web from the phone's own browser, even post on LinkedIn. One phone number, many agents.

Why not the official APIs? For small Italian businesses WhatsApp *is* the CRM, and the official Business API was too slow and heavy for what they needed. So we built a hub that exposes the phone as tools (MCP + REST) and lets any agent book it.

What actually broke, in order:

1. **The UI changed under us.** WhatsApp moved the "invite via link" and "add member" buttons in a single update. Every group action failed for a day until we patched the accessibility selectors. If you automate UIs, budget for this as a recurring cost, not a one-off.

2. **Shared inbox = ownership problem.** Five agents, one number. When a customer replies, whose conversation is it? We ended up tagging every inbound message with an "owner" (whoever last wrote to that number) and each agent filters for itself. Getting this wrong means agent A answering agent B's customer, with the wrong name in the greeting.

3. **Retries create double answers.** Our webhook retries 3 times. One project had an agent reply twice, 7 seconds apart, with two different greetings. The fix was on their side: an atomic lock keyed on the message id before generating a reply. Idempotency matters more than smart prompts.

4. **Rate limits come from the platform, not from you.** We added a daily cap on *first contacts* (people the number never talked to) because that's what gets numbers restricted, not total volume. Replies to people who wrote first are basically free.

5. **One screen, many agents.** The phone has one screen, so screen actions need a lease (book, act, release, max 10 minutes). Without it two agents typed into the same chat. Headless browser sessions on virtual displays solved most of it, since they don't need the physical screen at all.

What I'd tell anyone trying this: the LLM part was the easy 20%. The other 80% is boring plumbing: ownership, idempotency, leases, and noticing when the UI changed.

Curious if others here run agents on real devices instead of APIs. How do you deal with UI drift? Screenshots + vision, accessibility trees, or something else?
```



document fetched 2026-11-01 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1xk2uqp/we_gave_our_ai_agents_a_real_android_phone
