Valentine Tikhomirov

React Native engineer, 9 years in mobile. Exploring how AI agents change the way we build apps, and writing down what works. Blog:

[Valentine Tikhomirov @zemant91]

@blazejkustra_ Have you crossed over to the dark side?)

[Valentine Tikhomirov @zemant91]

@thymikee It’s sad for me.
I would like to, but I can’t 🥲

[Valentine Tikhomirov @zemant91]

@nativecookdev Good question. R1 had no gesture glitches. The closest visual bug was a button that ignored taps, and every setup fixed it fast, even shell alone. The hardest bug for all four was persisted state: a cart restored from storage. Gestures and animations are on my list for R2.

[Valentine Tikhomirov @zemant91]

1/ Does a device tool make a coding agent better at React Native bugs that only show up at runtime?
We ran Claude Code (Sonnet 5.5) on 5 such bugs with a bare shell, agent-device, Argent and Parallax.

Every setup fixed every bug. What differed was the cost.

[Valentine Tikhomirov @zemant91]

@blazejkustra_ @thymikee @agent_device Thanks, Błażej! Your notes made the harness better for every setup, not just Argent: skills and subagents on, warm servers, cold start as its own column, tokens split. Rerun of all four coming.

[Valentine Tikhomirov @zemant91]

@blazejkustra_ You were right, thank you. Our harness hid the simulator from Argent's server, so every Argent tap failed. On us, not Argent. The page now marks those runs as not valid, and all four setups will be rerun on a fixed harness.

[Valentine Tikhomirov @zemant91]

Correction: Argent's numbers in this round aren't valid. Our harness isolated each agent's home folder, which hid the simulator from Argent's server, so its taps failed. Thanks to @blazejkustra for finding it. All four setups will be rerun on a fixed harness

[Valentine Tikhomirov @zemant91]

@blazejkustra_ Thanks, Błażej! DMing you now. Happy to rerun Argent your way and update the page.

[Valentine Tikhomirov @zemant91]

@thymikee @agent_device Makes sense: the agent gets a settled screen in one answer instead of asking again. A good lesson for us too.

[Valentine Tikhomirov @zemant91]

@thymikee Agreed, and it cuts both ways: the bench found a gap in ours too. Parallax couldn't read app storage without the SDK, fixed in beta.3. I'll rerun on agent-device's next release and share the results.

[Valentine Tikhomirov @zemant91]

5/Method, limits and raw runs: https://t.co/WFUe5Pyxs8 Code, app, bugs and raw runs: https://t.co/ipIoYAox2K
Thanks to @thymikee for reviewing the method before we published. Two of his notes are in it: every run starts cold, and cost per fix is a mean.

[Valentine Tikhomirov @zemant91]

4/ It found a gap in Parallax too. Without our SDK, it couldn't read the app's storage, so the agent tapped through screens.
beta.3 reads storage, and that bug went from 92 s to 63 s. beta.4 added a navigate command: 57 s (3 runs).

[Valentine Tikhomirov @zemant91]

@progerallworld It already does: the same answer opens with what was tried ("Tapped count+ at (x, y)… no change"), then what's on screen; a missing id names the id it looked for. Typing reports only the length on purpose, since the text may be a secret. Thanks for reading this closely!

[Valentine Tikhomirov @zemant91]

1/ Parallax beta.10: when the agent's action misses, the answer now shows what was on the screen.

Also: a click mark in the window, and faster swipes on a real iPhone.

[Valentine Tikhomirov @zemant91]

5/ Limits: in auto mode Claude asks nothing, so there is no card. Edits made through the shell show in the changes, not as a card.

The desktop app is in closed beta. Want an invite? https://t.co/kFa2jN5tLu

[Valentine Tikhomirov @zemant91]

1/ Parallax beta.11: see the agent's edit before you approve it.

When Claude Code or Codex asks to edit a file, the card shows +n −m and the new Code tab shows the diff. Allow, and Fast Refresh shows the result on the screen.

[Valentine Tikhomirov @zemant91]

4/ You can't approve an edit whose diff isn't on screen. The edit stays in the window's memory until you answer, and secrets are masked.

After it: "Ask the agent to undo" or "Comment on this change" fills in a message. The window never writes your project.