Make the mobile agent prove the screen, not just write it

Inputs

• A screen spec with loading, empty, error, and offline states
• Theme tokens and accessibility requirements
• One realistic fixture, not placeholder lorem ipsum

Agent loop

1. Generate the smallest widget tree
2. Add a golden test for light and dark themes
3. Add semantics checks for icon-only actions and form fields
4. Run one device smoke test for navigation and keyboard behavior
5. Capture the screen states for review before opening the PR

Review question

Can a person understand what to do when the network is slow, the list is empty, or the action fails?

Research note

Adapted from the broader agent workflow pattern of pairing implementation with tests and visual proof, discussed by @kaxil: https://x.com/kaxil/status/2037503513350005134