Solo dev builds autonomous Android agent with 30 tools, on-device Gemma 3 1B, and a hard policy gate
Adventurous-Win6029 · reddit · 2026-09-06
A solo developer built a fully autonomous Android agent in native Kotlin that controls the phone via the accessibility tree, running a full perceive-think-act-verify loop.
Key design:
- 30 tools: flashlight, clipboard, SMS, app launch, web search, navigation in any app; spoken requests get spoken answers even with the app backgrounded.
- Two-tier brain: an on-device Gemma 3 1B handles simple commands (battery, flashlight, opening apps) with zero latency; harder tasks escalate to a 70B cloud model. The router is deterministic — keyword match + intent slots, no model call to decide routing.
- Policy gate: every tool call passes a manifest-checked gate that blocks flagged actions rather than log-and-allow, with operator taps resolving anything flagged. The architecture was ported from a project measuring prompt injection defenses across 112 models.
Other numbers: 134 unit tests, 8/8 device regression, 8.7 MB release APK (R8 minified), named recipes for one-shot teaching and replay of multi-step tasks, and deep perception reading structured data across 20 screens (e.g., product + price from Amazon searches).
The author's core argument: autonomy without a hard gate isn't autonomy, it's a liability — an agent that can do anything the phone can do needs a layer that can say no.
More from coding & agent
- 105 Real Bugs Tested: GPT-6.1 Sol Over 10x Cheaper Than GPT-5.6 Sol — PawelHuryn · 2026-10-23
- Dev builds SkiFree-style sled game in one prompt, praises DeepSeek Flash — victormustar · 2026-10-23
- Every ditched personal agents for a shared Claude Managed Agents coworker in one week — every · 2026-10-23
- WorkOS shows how to let AI agents safely sign up for your app with auth.md and ID-JAG — AI Engineer · 2026-10-23
- Devs on Reddit swap tactics for stopping prompt regressions in AI agents — SleepFlimsy6963 · 2026-10-23
- Claude Code + ComfyUI MCP workflow for client work: paid models or open weights? — newaccount47 · 2026-10-23