Agents write the code, phones approve the merge: device-in-the-loop CI for AI-built mobile apps

See how AI agents write mobile app code and real phones approve merges, with CI and device screenshots ensuring quality and agent loop survival.

Claude Code Gemini GitHub Actions Tmux MCP

Overview

An AI-assisted delivery workflow where coding agents ship mobile apps and physical phones give the final approval. A GitHub issue gets claimed by a Claude Code agent running in a long-lived tmux loop. The agent opens a PR with tests. CI on a self-hosted runner then installs the build on a real Pixel and a Wear OS watch through custom MCP device-control servers, runs the acceptance check, and posts device screenshots back to the PR. Live: the loop end to end on real hardware, MCP tool-call traces, the CI wiring, and the watchdog code that keeps a multi-day agent loop alive. Prerecorded backup for the device segment in case venue wifi dies.

Video

Transcript

Generated 3 days ago

Summary

The speaker presented an autonomous app development system, addressing the challenge of building apps with minimal human interaction. The demo showcased a fitness app mirrored from a Pixel Watch, highlighting the underlying infrastructure. The system leverages an open-source AI module from OpenAI, operating with a strategic room that uses a reactor level, objective graph, and program-based task creation. These tasks are converted into GitHub issues, resolved using CLI tools and Gemini. Pull requests are then reviewed by a separate AI, which also handles CI/CD checks, including on-device verification, to ensure robust and autonomous app development.

View full transcript

Hello, ma'am. I'm Andrey. I'm doing some exercise because I've been sitting and, filming this light. So here I'm mirroring my watch. It's a Pixel watch, and I'm showing the app I do with the eye with very little interactions for me because I decided to do something for my family. So here you go. I got I got my samples and workout, and I'm putting it to the board. But the point of this talk is not to steal a demo, but actually talk about the all the infrastructure that goes to build autonomous apps. So this visualization is OpenAI, open source AI. A module. So here I'm highlighting the module for that. After this demo. And the way it works, I have a strategic room which works on the reactor level with the programs, objective Graph, tasks tasks get created into GitHub issues. Here's an issue Local. Just using available tools using CLI and the Gemini, what you have to do with CLI. I tried factories. Easiest to do this. Let it, lose the capital and then CLI to build build the machine. So from here, we get the full requests. And new, a different reviewer mode reviews the full request and goes back and forth with the producer. I deal with all the CI, CCCI, checks, including on device check. Here's an example of a verification test for another app, a simple buffer level. And, after we have, more complex, we should require more complex assessing investment devices. I use, available I use homemade MCCs to manage And we're gonna move another level. So what I'm demonstrating Singh, a more complex test that so it's a public level application. And, I have a motorized I have a robotic gun on my AI, which the test would be something like this. What if you level, this is port? Will it actually come, to level position? Was not to have a recursion and also to pass the superiority tests up to this goal? Any questions? Yes? The the AI like, your organization of packages into CEO, like, can you show that a bit more? It is an open source project I use in, like

Links

https://lincinnovations.com/blog
LINC Innovations analyzes mobile CI, sensor tech, and device benchmarks.

Tech stack

Claude Code

Anthropic's agentic coding tool: Unleash Claude's raw power directly in your terminal or IDE to turn complex, hours-long workflows into a single command.

Claude Code is Anthropic’s powerful agentic coding assistant, designed for high-velocity development. It operates natively within your terminal, IDE (VS Code, JetBrains), or via a web interface, allowing you to delegate complex tasks like feature building, bug fixing, and codebase navigation. The agent plans, edits files, executes commands, and creates commits, maintaining awareness of your entire project structure. Internally, Anthropic engineers using Claude Code reported a 67% increase in productivity, demonstrating its capacity to deliver significant gains for Pro and Max plan users.

https://claude.com/code

View projects
Gemini

Google's natively multimodal AI model: understands and operates across text, code, audio, image, and video.

Gemini is Google's most capable and general AI model, engineered from the ground up to be natively multimodal: it seamlessly understands and combines information across text, code, audio, image, and video inputs. The technology is optimized for flexibility, running efficiently on everything from data centers to mobile devices. It is deployed in three key sizes: Ultra (for highly complex tasks), Pro (for broad scaling), and Nano (for efficient on-device tasks). Developers access this power via the Gemini API to build next-generation applications.

https://deepmind.google/technologies/gemini/

View projects
GitHub Actions

GitHub Actions automates your software workflows (CI/CD, testing, deployment) directly within your repository.

GitHub Actions is your native, event-driven automation platform: it executes custom workflows right inside your GitHub repository. Define multi-step jobs using YAML files, which must be stored in the `.github/workflows` directory. These workflows trigger on specific events, like a `push` to a branch or a `pull_request` opening. The platform supports world-class Continuous Integration and Continuous Deployment (CI/CD), letting you build, test, and deploy code across any language (e.g., Python, Node.js, Java). Runners—Linux, Windows, or macOS virtual machines—execute the jobs. For security, confidential data is managed via GitHub `secrets`, ensuring credentials stay out of your code.

https://github.com/features/actions

View projects
Tmux

Tmux is a terminal multiplexer: it allows users to manage multiple terminal sessions from a single window, with the critical ability to detach and reattach.

This technology is a game-changer for terminal power users. Tmux operates as a terminal multiplexer, enabling you to run multiple programs within a single interface, managing them across distinct sessions, windows, and panes. Its core value is persistence: you can start a long-running process, detach the session (using the default prefix key, `Ctrl+B`, then `D`), close your SSH connection, and reattach later from any machine to find the job still running. Furthermore, it facilitates organization by allowing a single window to be split into multiple rectangular panes (e.g., horizontally with `Ctrl+B %` or vertically with `Ctrl+B "`), effectively acting as a window manager for your command line.

https://tmux.github.io

View projects
MCP

MCP is the open-source standard for securely connecting AI agents (like LLMs) to external tools, data, and enterprise workflows.

The Model Context Protocol (MCP) functions as a standardized integration layer: think of it as a USB-C port for AI applications. Developed and open-sourced by Anthropic, this protocol allows large language models (LLMs) to access real-time context and execute actions via external tools like GitHub, Jira, or proprietary databases . It uses a simple JSON-RPC interface to define tools, schemas, and endpoints, which enables AI agents to perform complex, state-changing tasks—such as creating a GitHub issue or running a test script—rather than just generating text . MCP is essential for building agentic AI systems that can autonomously pursue goals and operate within defined safety and permission boundaries .

https://modelcontextprotocol.io/

View projects

Agents write the code, phones approve the merge: device-in-the-loop CI for AI-built mobile apps

Related talks