Feedback: Codename One and AI agents are now a remarkably strong integration for real app development

[jsfan3]

Recently Shai documented an important change in the Codename One blog: new Java 17 projects generated with the Initializr now include AGENTS.md and a vendor-neutral Codename One authoring skill, so modern generative AI tools and AI agents can understand what Codename One is and how to work with it:

https://www.codenameone.com/blog/skills-java17-and-theme-accents/

Between yesterday and today I ran a practical experiment to test what is actually feasible with the current state of the art.

I do not want to discuss here the ethical and labor-related issues around generative AI, because this is not the right place for that discussion. I have already written about that extensively on my own blog, including in dialogue with the engineer Giulio Ripa:

https://www.informatica-libera.net/

Here I only want to report a technical result. I also sincerely hope this post will be indexed by search engines, because I think it says something useful about the current relationship between Codename One and AI-assisted software development.

The experiment was this: I imagined a small but real Android app whose development would include several non-trivial complications.

The app is called StreakApp. It tracks language-study streaks by reading Android foreground app usage for apps such as AnkiDroid, Duolingo, Drops, Rosetta Stone, Talkpal, and Quizlet.

The requirements intentionally included:

• A new Codename One Java 17 project generated with the Initializr.
• Use of recent Codename One APIs and conventions, as documented in the Codename One blog. This is important because the AI agent had to work with very recent information.
• Native interfaces. The AI had to distinguish between Codename One APIs and native Android APIs, then bridge them correctly.
• I did not specify which native Android APIs were needed, nor did I tell the AI which Codename One build hints were required. I let the AI infer this by itself. It correctly had to deal with things such as Android usage statistics, app launching, android.manifest.queries, ProGuard keep rules, and permissions/build hints.
• A detailed description of what the app had to do, but without implementation details and without explaining Codename One best practices.
• A general verbal description of the layout and visual style, leaving the concrete UI decisions to the AI.
• English and Italian localization.
• Light and dark themes.
• Text-size handling from small to very large values, without breaking the layout. This is not trivial in a mobile UI.
• Real-device Android testing during development. Because native interfaces were involved, the simulator was not enough.
• Android APK generation through the Codename One build server using my Codename One account.
• Automatic installation and updating of the APK on my real Android device connected by USB.
• Visual testing on the real device, including screenshots, layout checks, color checks, and CSS/Java corrections based on those screenshots.
• Automatic preparation of GitHub documentation, screenshots, repository metadata, and an APK release, while removing sensitive local data such as the Android keystore and its password.

The result was very good.

Almost everything worked. The app was created, built, installed, and repeatedly tested by the AI agent during development, with very little manual intervention from me. The whole process took several hours. During that time I mostly supervised the work and occasionally performed manual checks on the device, reporting small issues or refinements back to the AI.

The app is now public here, with source code, documentation, screenshots, and an APK release:

https://github.com/jsfan3/streakapp

There was one bug the AI did not solve correctly by itself: it had misunderstood the difference between CN.getCurrentForm() and Lifecycle.getCurrentForm() in one place where live UI refresh mattered. I pointed this out manually, and after that the fix was straightforward.

I mention this detail deliberately. Human review was still necessary, and I think this is a good thing. It means that the human developer is still part of the software development process, at least for now. Future AI models will probably be less easily confused by methods with the same name in different classes, but today this kind of review still matters.

My final verdict from this experiment is:

Codename One is a mature, stable, and complete Java cross-platform framework for modern real-world app development with generative AI, AI agents, and a test-driven workflow.

This setup can relieve the developer from having to manually handle every complex implementation detail, such as native interfaces and related build hints, while still allowing the developer to keep a higher-level architectural and product view of the work.

In my opinion, this is perfectly aligned with the current transformation of software development caused by generative AI.

In the next few days I will use StreakApp in real life. If I do not notice anything else that needs refinement, I may also publish it on Google Play.

Two final notes.

First, I have already tested Codename One + generative AI / AI agents on a more complex project, which I plan to publish later as a tutorial. In that case too, I observed that the AI understood good Java development practices, split classes into appropriate packages, and produced code that was reasonably organized and easy to inspect.

Second, for this specific experiment I used IntelliJ IDEA Ultimate with AI Chat configured as:

• OpenAI Codex
• Agent mode with full access
• GPT-5.5, reasoning effort xhigh

This is only one possible setup. The important point is that Codename One is now well positioned for integration with the main generative AI tools and AI agents.

May 26, 2026

[shai-almog]

This pretty much mirrors my experience with the current crop of AI agents. I love them because they are so good with the easy/tedious stuff and make the REALLY hard stuff that requires long iterations/patience possible.

Thanks to Java's strict syntax and deep documentation coupled with the very fast test cycles we can provide in the simulator, agents can be far more productive with Codename One than with any other platform. IMHO this is a good thing. As a small company it was nearly impossible for us to keep up with the likes of Google as a competitor, now we're aiming to surpass them. I think smaller/individual developers now have a chance to compete against major corporations on their own level.