On the Adoption of AI Coding Agents in Open-source Android and iOS Development
This paper presents the first empirical study of AI coding agents in open-source mobile development, analyzing 2,901 pull requests to reveal that Android projects receive more AI contributions with higher acceptance rates than iOS, while routine tasks are more successful than structural changes.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine the world of software development as a massive, bustling construction site. For years, humans have been the architects and builders, laying bricks (code) and fixing leaks (bugs). But recently, a new crew has arrived: AI Coding Agents. Think of these agents as super-fast, tireless robot apprentices who can write code, fix errors, and suggest improvements in seconds.
This paper is like a site inspector's report on how these robot apprentices are doing specifically on two major types of construction sites: Android (the open, customizable city) and iOS (the strict, curated garden).
Here is the breakdown of their findings, using simple analogies:
1. The Setup: Two Different Neighborhoods
The researchers looked at 2,901 "work orders" (called Pull Requests or PRs) that these AI robots submitted to open-source mobile apps. They compared how these orders were treated in two different neighborhoods:
- The Android Neighborhood: A place with many different building styles, open rules, and a lot of variety.
- The iOS Neighborhood: A place with strict building codes, uniform designs, and very specific rules.
2. The Big Question: Who Gets Hired? (Acceptance Rates)
The researchers asked: When the robot apprentice submits a fix or a new feature, do the human bosses actually accept it?
- Android (The Open City): The robots got hired more often here (71% acceptance). However, it depends heavily on which robot you use. Some robots (like Codex) are like star students who get hired almost every time (77%), while others (like Copilot) are like new interns who get rejected more often (28%). The bosses here are picky about who they hire.
- iOS (The Strict Garden): The robots got hired less often here (63%). But here's the twist: it didn't matter which robot they used. The bosses were uniformly cautious. Whether it was the star student or the new intern, the acceptance rate was roughly the same. The strict building codes of iOS make everyone nervous about letting robots in.
The Takeaway: If you are building an Android app, the choice of AI tool matters a lot. If you are building for iOS, the AI tools are all treated with the same level of skepticism.
3. What Kind of Work Gets Done? (Task Categories)
The researchers also looked at what the robots were trying to do. They found a clear pattern:
- The "Easy Jobs" (Routine Tasks): Things like fixing a typo, changing a button color, or translating text (Localization).
- Analogy: This is like the robot being asked to "paint the fence" or "mow the lawn."
- Result: These were accepted very quickly and often, especially on Android.
- The "Hard Jobs" (Structural Changes): Things like rebuilding the foundation, changing the engine, or reorganizing the whole house (Refactoring/Build).
- Analogy: This is like the robot trying to "move the kitchen" or "reinforce the roof."
- Result: Humans were much slower to accept these. They took longer to review and were rejected more often because they felt risky.
The Takeaway: AI is great at the small, routine chores. But when it tries to do major structural changes, human bosses hit the brakes and take a long time to think it over.
4. How Fast Did They Finish? (Resolution Time)
- iOS Speed: Surprisingly, even though fewer iOS requests were accepted, the ones that were accepted got processed 18 times faster than Android. It's like a strict but efficient manager who says "No" quickly, but if they say "Yes," they get it done in a flash.
- Android Speed: Android was slower overall, but it showed a weird trend. In the middle of 2025, things got faster (the robots and humans learned to work together), but then by late 2025, things slowed down again. It was a rollercoaster of efficiency.
5. Why Did Some Get Rejected?
The researchers peeked at the rejected requests and found some funny reasons why robots get fired:
- The "Rebase" Problem: The robot tried to build on top of a wall that the human had already moved.
- The "Silent Treatment": The robot submitted the work, but no human reviewed it, so it expired.
- The "Broken Pipeline": The robot's work failed the automatic safety checks (like a car failing a safety inspection).
The Final Verdict
This paper tells us that AI coding agents are real, but they aren't magic.
- They are best at small, routine tasks (fixing bugs, changing colors).
- They are struggling with big, structural changes (re-architecting code).
- Android is more willing to try different robots, while iOS is more cautious and uniform.
- Humans are still the bosses. We need to choose the right robot for the job, and we need to be ready to step in and fix things when the robot tries to do something too complex.
In short: The robot apprentices are helpful, but they still need a human foreman to look over their shoulder, especially when they try to move the walls.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.