# Built a fully autonomous Android agent — 30 tools, on-device model, and a policy gate that blocks every
  untraced action — retained record

> Retained reddit record from reddit:AI_Agents.

Canonical: https://highsignal.app/data/records/745d653e1ea25bc1

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

  untraced action

source: [reddit](https://highsignal.app/data/reddit) (reddit:AI_Agents)  published 2026-09-05 16:02 ingested 2026-09-06 02:36

[original source ↗](https://www.reddit.com/r/AI_Agents/comments/1w84mmf/built_a_fully_autonomous_android_agent_30_tools/)

## retained text



```
I've been building an autonomous Android agent that controls the phone through the accessibility tree. It reads the

  screen, decides what to do, does it, then verifies it worked — a full perceive-think-act-verify loop.



  What it does:



  30 tools — flashlight, clipboard, SMS, app launch, web search, navigation through any app. A spoken request through

  the side button gets a spoken answer back, even with the app backgrounded.



  The brain is a two-tier system: a Gemma 3 1B model runs on-device for simple commands (battery, toggle flashlight,

  open an app) with zero latency. Anything harder escalates to a 70B cloud model. The router is deterministic — keyword

  match + intent slot, no model call to decide which model to call.



  The part I think is worth discussing here — the policy gate:



  Every tool call passes through a manifest-checked gate before it touches the device. The gate doesn't log-and-allow —

  it blocks. An operator tap resolves anything the gate flags. I ported this from a separate project where I was

  measuring prompt injection defenses across 112 models. Same architecture, enforced on-device.



  This is the part I'd push back on with most agent frameworks: autonomy without a hard gate isn't autonomy, it's a

  liability. The agent can do anything the phone can do — which is exactly why it needs a layer that can say no.



  Numbers:



  \- 134 unit tests, 8/8 device regression suite

  \- Release APK is \~8.7 MB (R8 minified)

  \- Hands-free assist works with the app backgrounded

  \- Named recipes: teach a multi-step task once, replay it later

  \- Deep perception reads structured data from apps (e.g., product + price from an Amazon search across 20 screens)



  Native Kotlin, built solo.



  Happy to go deeper on the gate architecture or the tool loop.


```



document fetched 2026-09-06 00:18 · canonical https://reddit.com/r/AI_Agents/comments/1w84mmf/built_a_fully_autonomous_android_agent_30_tools
