May 20, 2026
Teaching My Phone-Based AI Agent to Fix Its Own Mistakes
Error recovery, file persistence, and fallback planning for Julian: a self-hosted assistant on a Galaxy S21.
Julian runs on a Galaxy S21 I already owned. No subscription. No cloud that quietly becomes the product. That part was the point — and also the trap. Local agents fail in ways demos never show: a half-written file, a tool that hangs, a plan that sounded fine until the phone got warm.
This essay is about teaching that agent to recover. Error paths, persistence that survives a reboot, and fallbacks that don’t pretend the first try always works. The interesting work wasn’t the model. It was the boring scaffolding that keeps a personal agent from lying to you when it breaks.
If you’ve ever watched a local assistant stall mid-task and shrug, this is the missing half of the build log.
That's the setup. Continue on Medium
Follow the work
New essays by email: a list I own, not a Medium follow.