ieve's Squall of Consciousness

Entry 16 - The Promise of Agentic Coding

(April 11, 2026)

The classic approach was fun

I have liked programming since ... always. In the early 2000's, in middle school, I was creating Geocities sites, assembling wee javascript codes to play music based on dropdown selection, hoarding cool gifs and divider images. Around 2018 I picked up Ruby, and then Rails, making web apps to do silly things like automatically send SMS Cat Facts to my friends, or track things in CRUD apps.

Of course when it came time to stop screwing around and go to school (finally, at around 22 years old), I decided ... ya know, I can already do a lot of this stuff on my own, it would be a waste of time to have school teach me, and so I did mechanical engineering instead of computer science / engineering. (I would probably make a lot more money if I kept to the coding route and moved to the Bay... but... it was not my path.)

I'm 35 now so... I've been doing this a long time, and I completely subscribe to the joy that comes from flow state noodling over a thorny problem for hours or days. Booting up a REPL or dropping into a debugger and exploring the state and what exactly does this function return and what's the data format and what do I do with it next... it's great fun and very rewarding.

But for me, the horizon has flattened

I suffer from (likely) untreated ADHD, or at least a shattered attention span from existing in the modern world, and there's an activation energy that meant my hard drives and github are scattered with the bones of a hundred half-baked projects that amounted to nothing.

I have found LLMs lifechanging in this respect - I'm able to take on and make progress on projects that I otherwise would have stalled on by being unable to get over the activation energy barrier.

Like my frost mage costume I made in December ... sure I had a couple Youtube video guides on how to sew a traveler's cloak, but being able to describe my frustration and what I'm seeing with the particular stitch I'm trying to use. By being able to tell Google Gemini "please get the transcript from this Youtube video" and then discuss the details with it, to understand what to do in the gaps where the original author didn't explain things clearly enough for me.

Or by being able to ask for a simple program for the microcontroller to achieve what I wanted. Yes I love programming, but figuring out the syntax, figuring out the basic logick, while sewing a cloak, while soldering my circuit together, while sanding a staff, while styling the staff with hot glue icicles and affixing my LED strings.

Yes I'm technically capable of undertaking such a project. I can do hardware, I can do software, I can rub sandpaper and a rasp over a stick I found in the canyon, I can shop for electronics.

Could I actually wrangle all of the cats in my brain and push a project like this to completion without an LLM? Probably not. Just me personally, it's a no.

A progression of coding with LLMs

So of course I code with LLMs. I'm not a software engineer, most of my brain cycles in the week go toward medical device development and sustaining tasks, so... I don't feel too bad about leaning on LLMs for help with understanding patterns, language features, best practices for how to accomplish certain things.

I started with asking the chatbots questions and getting code suggestions in response. I (still) believe that actually writing things out is a way to learn and understand versus copy pasting, so I (still often) bring the chatbot window side by side with my code editor and type in the code myself. This lets me think about things as the characters are going in and make decisions as I go.

But of course Claude Code was the new hotness so I bought some Anthropic API credit and tried it out. It's magick. (We now have the leaked source, and I dabble in creating my own LLM harnesses for local use, so I understand that it's not magick, and I will still declare it to be magick.) Especially with a compiled, statically typed language, and with consistent use of automated unit, integration and (if possible) system testing, the feedback loop is so tight that overall success rate of code that _technically works_ is essentially 100%.

(Whether you're getting the right features built, and whether the codebase is becoming bloated with duplicate / overlapping features and approaches are totally separate questions.)

I think Claude Code is best suited to interactive development, considering the landscape of tools available now. Each diff that pops up lets me update my mental model of the codebase in real time, and lets me steer the development if the agent didn't understand my intent, or I see that I want things to be differently.

But... those Anthropic tokens add up quickly. I wrote last month about vibe coding a command line fitness tracker for myself for about $20 in tokens:

Strong Vibes (Mar 24, 2026)

I'm a heavy user so I already pay Google for a pro plan, and saw that my plan includes 100 Jules requests per day with a max of 15 concurrent tasks. (This is ... way more than I could ever use.)

(Web link) Google Jules

Jules is a very different style of coding agent. Instead of running in my local environment, it creates an ephemeral sandbox for each instance of the tool.

An example task that I passed to Jules before going to bed last night:

Evaluate {main file of codebase} for variables that should be included in the json config.cr import/export

For example, ollama parameters should be user configurable. Selection of target and maximum for context and memory windows should be configurable.

Jules then launches a sandbox (which is cloned from some saved state as a starting point with compiler and any other essential tools installed), clones repository HEAD from Github, creates a plan to accomplish the task, asks follow up questions if needed (or if requested in planning mode), and goes off to do the work. It runs tests to make sure things pass, and then when everything appears ready, opens up a pull request in Github for me to review and merge into HEAD or reject.

If I want it to make any changes, I can just tell Jules what to change / what to add and let it go off to do its work again.

Unprofessional Cat Herder

So in the dream version of how this all works, I tell Jules to implement a bunch of features, I go to sleep or for a walk or work on some other part of the codebase in the background, come back to some pull requests, and get a bunch of fixes / feature adds with zero effort.

...

But in reality, my request to change my hardcoded module constants configuration (which needed me to recompile on any config change) into a JSON-backed configuration file was started in parallel with a half dozen other changes. (Jules scans codebase and suggests changes for things like performance improvements, code readability / health, security issues, or just other changes I thought of at the same time.)

So a half dozen dutiful little workers cloned the same HEAD that the first one started with and their code changes often either interact with the configuration in some way, or they 'helpfully' made some little formatting change while they were at it that touched a configuration line.

So most pull requests involve resolving merge conflicts. In some cases things are so tangled up that it's not even worth accepting the pull request - either for a small suggested change that I didn't even care that much about, or something that would be easier to just make the edits manually in my local environment.

...

Google Jules describes itself as a tool that "does coding tasks you don't want to do." I would say that promise has not been fulfilled.

I want to figure out how to automatically migrate my configuration file when I add a new field to the schema and have something automatically coordinate & merge parallel improvement tasks.

Instead, I have a tool that could effortlessly write the code for configuration migration, but I'm left playing as Git Janitor and Unpaid Release Engineer.

It's a real "wait I wanted the robots to fold my laundry so I could spend the day doing art" vibe.

- 🜌

Back to Squall of Consciousness Index