Developer demonstrates cross-app AI Agent

Article

iPhone version of Doubao Phone Assistant appears? Developer demonstrates using Opus 5.5 Operating iPhone

Beating AI Breaking news: developer Finn Voorhees demonstrated an AI Agent. He said just one sentence, asking the Agent to find the cheapest Ryanair flight to London departing the next day. The Agent then opened Ryanair, entered the destination, selected the date, checked the fare, and finally returned to the chat interface to report the result. No computer was involved at any point.

The Agent runs in PiChat, an iPhone chat app developed by Finn. Under the hood, it uses his previously open-sourced PiDurableKit and Apple's JavaScriptCore to run the Pi Agent on the phone, handling model calls, tool execution, and task management. The video uses Claude Opus 5.5, while model inference still takes place in the cloud. To let the Agent continue working after switching to another app, PiChat also uses Apple's background task mechanism and SQLite to save execution progress, making it possible to resume after an interruption.

The most critical cross-app control code has not yet been published. When iPhone Control is enabled in the video, the system requires the lock-screen passcode to enable XCTest UI automation. XCTest is Apple's UI testing framework and can simulate taps, swipes, and text input. The open-source TouchSynthesis project uses XCTest's private interfaces to generate touch events on the iPhone itself, without requiring real-time computer control. However, Finn has not confirmed whether this approach was used.

Original link https://m.theblockbeats.info/flash/371032

Key points

Finn Voorhees demonstrated an AI Agent。
The Agent searched for the cheapest Ryanair flight to London departing the next day, then independently opened Ryanair, entered the destination, selected the date, and checked the fare.
The Agent runs in PiChat, an iPhone chat app developed by Finn Voorhees.
PiChat uses PiDurableKit and JavaScriptCore to run the Pi Agent on the phone.
The video uses Claude Opus 5.5, while model inference still takes place in the cloud.
PiChat uses Apple's background task mechanism and SQLite to save execution progress.
The cross-app control code has not been published.
When iPhone Control is enabled in the video, the system requires the lock-screen passcode to enable XCTest UI automation.
TouchSynthesis uses XCTest's private interfaces to generate touch events, but Finn Voorhees has not confirmed whether this approach was used.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

This content describes a demonstration: with just one sentence from the user, an Agent on the phone can open another app to search for flights, then return to the chat interface and report the result. The Agent's model inference takes place in the cloud, while the phone runs the task and saves its progress.

Why it matters to readers

It demonstrates AI Agent the possibility of executing cross-app tasks on an iPhone, but the key control code has not been published, and the specific implementation requires further confirmation.

It can be understood as a chat assistant that can operate phone apps on the user's behalf; the current demonstration does not mean that ordinary users can already use it reliably.

Risks and unknowns

PiChat’s cross-App control code has not yet been made public.
Finn Voorhees has not confirmed whether TouchSynthesis or private XCTest interfaces are used.
The evidence alone cannot confirm whether the demonstration can be reproduced by ordinary users.
The evidence does not clearly indicate whether the “Doubao Phone Assistant for iPhone” mentioned in the title is related to the demonstration.

Developer demonstrates cross-app AI Agent

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center