ProgrammablePhones

ProgrammablePhones

Created Thursday 14 May 2026

GridCalc in its current layout.

The Phone Becomes Programmable

Over the past several months it's become clear to me that a revolution in mobile personal computing has begun, spurred by recent advances in AI model capability and coding agents.

Until around December, most programmers I knew, myself included, had plenty of ideas for mobile apps to make life easier. But few had time to invest in the ecosystem: learning frameworks, platforms, and languages in order to ship either iOS or Android apps with an amount of effort that made sense compared to other obligations or pastimes.

I've always been a little frustrated by the fact that I carried a powerful computer on my person at almost all times, and yet I couldn't easily write software for it.

That's changing. The device in my pocket is starting to feel less like a sealed appliance and more like a personal computer again: something I can shape around my own habits, sensors, workflows, and annoyances.

The App I Actually Wanted

Things started changing for me when GPT-5.2 was released on December 11, 2025. Shortly after I finally gave a terminal AI coding agent, OpenAI's Codex, a fair shake, and was wildly productive with Rails.

Then in January of 2026 I had an idea for the kind of iOS calculator I wanted, and went from a UI sketch on a sticky note to a release in the App Store 9 days later: GridCalc, an RPN spreadsheet with all sorts of whizbang features. I've used it every day since.

I Wasn't Alone

I'm not the only one going nuts building stuff for mobile. Last month it was reported that the App Store saw an 84% surge in new apps in Q1 2026.

My friend Conrad Barski built a series of utility apps for himself on Android, including one that takes input from a smart ring.

Chris Meiklejohn has also been publicly building a mobile app and sharing a lot of good work around agentic coding research and his own experiences on his blog (strongly recommend).

Other friends built fitness apps for heart-rate-guided aerobic training.

Getting It Onto a Phone

There are slight differences in setup across platforms. Currently Android is a bit easier. Conrad builds his apps in the cloud on Expo. Some of my friends keep Android tools on their Macs, taking advantage of sideloading. I wanted to make a product in the App Store, so I installed Xcode locally and published GridCalc there (I even made a YouTube demo video in iMovie with substantial AI guidance and an ElevenLabs voice, but that's for another post).

But publishing is not the whole story. For AI-powered mobile personal computing, the important loop is simpler: build the utility you need, get it onto the device you use, and let it become part of your day.

I love it. The computers we physically carry, physically touch, and stare at all day are finally doing our bidding.

The Hard Part Is Still Hard

There are problems, of course, but as far as I can tell they are those inherent to AI coding generally. Currently, AI agents work well with established languages, toolkits, and platforms. But like any software, mobile apps are vulnerable to self-induced bloat that can eventually bring any software project, especially an aggressively AI-coded one, to a grinding halt, as state space and complexity exceed both the context window of the LLM and the time budget of even automated QA.

UI and interaction toolkits on mobile might be especially prone to this, due to added complexity from touch interaction, multiple-device support, and screen size and capability differences.

Ratcheting up your linters and build hygiene requirements is a way to mitigate bloat paralysis, but to make high-performance, reliable apps, you still have to know what you're doing, at least at a high level. When AI makes code cheaper, correctness and maintainability become the bottlenecks.

For lower-risk internal tools, AI already pays for itself in my workflows. Shipping mobile apps is different: higher risk, higher reward, and much less forgiving. Since I'm not writing the code for my mobile apps, but still staking a side business and my personal reputation on them, they have to be damn good.

What I'm Learning Now

For my part I'm leveling up on state machines, model checking, dependent types/theorem provers (again; currently looking at Lean), and statistical analysis. I used hierarchical state machines for GridCalc, and that thing has been rock solid. A lot of what Chris Meiklejohn wrote up in Getting Up to Speed on Multi-Agent Systems, Part 6: Verification Patterns connected with my recent experience and thinking around verification and QA.

I'm also taking (another) hard look at Backus's FL, his stab at a functional language over an algebra of programs (not types). I suspect there might be some total, closed programming paradigm like FL that offers the balance of power and restraint these AIs seem like they need to do a good job without skipping off the rails. I should be...