Revisiting AI Agents

Well, I'm back. AI has gotten better at coding since I last posted in September. Like, markedly better. Like scarily better. Like it's no longer a question of if AI can write an app from scratch, but more of why on Earth are we even bothering to type things with our human hands anymore?

As a human with hands (currently unbroken at time of writing), this concerns me. News like Block getting rid of 40% of their workforce because of AI really concerns unemployed me. It feels like those in the tech industry who don't start trying to ride the AI wave are going to be crushed by it. So, uh, guess it's time to grab my surfboard and start paddling. 🏄🏼‍♀️

The Initial Problem Statement

I have my AIPackingList app and I'm going to maintain the same boundaries as before - me no write code. My goal at the end of this is to implement a new feature that should require some restructuring of the existing code.

One big change between this and the last two segments is that I'll be utilizing Claude from Anthropic instead of Gemini from Google. The main motivation for this change is my husband offering to pay for the Pro version of Claude so I can really explore the depths of this AI rabbit hole without having to worry about immediately running out of tokens (please don't explain the concept of shared finances to me, it's a joke). I used Claude to rewrite the website that this post is hosted on and it went really well, so I am optimistic about this experiment.

Claude currently has a beta plugin for Android Studio / Jetbrains that appears to pretty much be a wrapper around its Claude Code terminal offering. I installed this, but ended up just using Claude Code in the terminal as they seemed functionally similar.

To begin with, I turned on Claude Code's "Plan Mode", which essentially turns the assistant into a read-only assessment mode, where you can consult it about a given codebase or ask it to develop a multi-step plan where you can then edit individual steps.

This was my initial prompt, which describes the new feature to add to the packing list app - templatization!

I want to add the ability to create a new list from a template. I'd like the app to have several templates available for packing lists for different vacations (ski trip, beach trip, weekend away, etc). The user should be able to add or remove items from these templates and save them for future use. These templates should be separate than the actual lists, which the user can add items to without altering whatever template it was based on.

 companion object {
     val DEFAULT_TEMPLATES = listOf(
         PackingListTemplate(id = 1, name = "Beach Trip", items = listOf(...)),
         PackingListTemplate(id = 2, name = "Ski Trip", items = listOf(...)),
         PackingListTemplate(id = 3, name = "Weekend Away", items = listOf(...)),
         PackingListTemplate(id = 4, name = "City Break", items = listOf(...)),
     )
 }

 init {
     viewModelScope.launch {
         dataStore.updateData { prefs ->
             if (prefs.templates.isEmpty()) prefs.copy(templates = DEFAULT_TEMPLATES) else
  prefs
         }
     }
 }

I love plan mode, at least just based on this initial usage. I can dig into the technical details or just skim them and accept edits to the code. This is very similar to shepherding a smart junior engineer - letting them run with it, but still checking that they're on the right track. I also love that it is just going to write unit tests without any additional prompting.

My only question now is can I change the default "auto-accept edits" prompt to "just send it".

Anyway, I let it spin for a bit and eventually it spit out the following:

● Build successful, no warnings. Here's a summary of everything that was implemented:

New files:

• TemplateItem.kt — data class for template items (id + text, no checked state)
• PackingListTemplate.kt — data class for templates
• TemplateItemRow.kt — simple item row with delete button, no checkbox
• TemplateCard.kt — card with in-place rename, item count, and "Use this template" button
• TemplatesScreen.kt — list of all templates with FAB to add new ones
• TemplateDetailScreen.kt — edit a template's name and items

Modified files:

• UserPreferences.kt — added templates field
• Screen.kt — added TemplatesScreen and TemplateDetailScreen screens + Saver support
• PackingListViewModel.kt — added init block to seed 4 default templates (Beach Trip, Ski Trip, Weekend Away, City Break), newListCreatedEvent SharedFlow, and 6 new template functions
• AllPackingListsScreen.kt — added Templates icon button in the top app bar
• MainActivity.kt — wired up all new screens and handlers

UX flow: Tap the list icon in the top bar → browse/edit templates → tap "Use this template" to create a new list with copied items → navigated directly to the new list. The original template is untouched.

✻ Cooked for 7m 17s

This is like... really good. It took care of the templating use cases and copying everything and... like the only thing I would want to add (maybe) is being able to delete a template? But honestly it took what would have taken me a bit to implement and verify and I probably would have missed a use case and just did it in 7m 17s. And you can turn on auto-accept so you can truly, truly let AI yolo you an app.

I kind of don't know where to go from here. This is the basic functionality I wanted when I set out with this app idea in my head. It's just... done - I just spoke a feature into existence in less time than it takes me to walk to a coffee shop and get a latte.

Cue the Existential Dread

Writing this conclusion has involved a lot of moodily staring out the window into the glum Seattle "spring". I set out in July asking if my job was obsolete as a tongue-in-cheek joke, and now I'm asking that as a serious question.

It's insane how well this technology has progressed from September, when I struggled to get agentic AI to write unit tests - let alone making reasonable code, making and verifying a plan, executing, test writing, and doing all that with minimal input from myself, a human-in-the-loop. I don't even have to copy/paste anything, besides the original prompt. And $20/month is significantly cheaper than my previous salary. It's hard to come out of this experiment and feel good about the future of humans in software engineering.

Granted, there's a lot more that goes into engineering than just typing a prompt into a terminal, especially when dealing with significantly large and complicated systems that hold more private data than a packing list. But for this example it was significantly easier to just make an AI do it than to contribute anything of note myself. I'd also argue that the spark of creativity - the idea to write the app is still firmly in the court of the human entity. But does that just make us all PMs to the robot underlings?

I don't know. My cat was pretty happy with it though, as he could sit in my lap and I could use my hands to pet him for seven minutes instead of typing for hours. I'd feel a lot better about it if I could earn money for taking care of my cat versus writing and designing code.

I think at this point this initial series of experiments can be called complete. I'm still going to be working on this, as this is still really only scratching the surface of AI in software engineering. But I think the question of "can AI code well enough to write features" has been answered.

Yes. Yes, it can.