I built native iOS + Android apps mostly by steering AI agents. What worked, what didn't

For the past few weeks I've been building a shared grocery list app, native on both iPhone (SwiftUI) and Android (Jetpack Compose), with a small Next.js backend. I wrote very little of the code myself: I mostly steered AI coding agents (Claude Code), one per platform, running in parallel.

What worked

1. Writing a short spec and a shared API contract before any code. Both agents build against the same contract, so the apps behave the same.

2. Asking every agent for screenshots from a real simulator or emulator, not just "tests pass".

3. Snapshot tests and strict CI as the safety net: they catch the "small" visual regressions agents love to make.

What didn't

1. Vague requests ("make it nicer") produce churn. Precise values (sizes, colors, offsets) produce parity.

2. Left alone, an agent will sometimes "fix" the test instead of the code.

3. iOS and Android drift apart unless shared constants live in one place.

I still make every product and design call, and I test on real phones every day. It feels less like coding and more like being a tech lead for a very fast, very literal team.

How do you work with agents? Do you let them touch UI, or only logic? And how do you review what they ship?