Leaning into agentic coding: (almost) a year in (short) review

As I look for work, please support me with a donation to the project development I'm spending in the Atmosphere — donate here

Last year, while I was marooned in France, I wrote a bit of a reflection on my stance of using generative AI to write software. I don't pretend - in this moment - like it's a bit of a double edged sword: I don't pay for things like Antrophic or OpenAI, given their explicit involvement in being used for violence. However, there's nothing stopping the eventual transition of large government agencies to newer open models built from the ground up from the millions of documents used for public policy and law - something that I wouldn't be surprised if it was underway. It would have to be open (at least a version of it on declassified information) since it's (the documents produced) funded publicly. But then again, we don't have federal or municipal telephony services nor Internet services. My approach to model work has been the following:

I can't afford the kind of hardware needed (currently) to run models at home, as most Americans who are living between jobs (like myself) can't afford things. And it'd be frankly wasteful for me to invest in that at the moment. However, my subscription pricing for the subscription above are - together - touching about $100 and I only move to another service if/when I hit the rate limit. Leaning on Deepseek makes this not a thing so I also switch to models like MiniMax and Z.AI when I need heavier reasoning and rely on Deepseek for "grunt work".

"You're ... Using ... AI"

Generative AI — yes. And frankly, it'd be a bit ignorant to not assume most of the things produced today isn't using it. This isn't defending the proliferation of it but more being aware of the state of affairs. I do believe a lot of discontent on the potentiality of it taking away jobs is a bit late in the tooth to address because folks have been losing jobs in tech routinely. I only see commentary about it now from some folks who have been generally quiet about things as a whole; indicating to me something of "I can only care when it's in my yard" mechanics of thought. I find it a bit exhausting but I'm also of the mind that software engineers are grossly overpaid and over privileged and that we need to crank it back and redistribute that wealth to folks who make companies go.

Back to the point at hand.

What's Your Stack?

I use OpenCode primarily as a multiplexer of model interactivity. It has a lot of plugins packed into it but the most notable has been the superpowers plugin as it puts in some sensible defaults. If I could get it in Pi, another tool I use, I would but it's not needed as I use Pi as my "slegehammer" for implementing plans. It not having a "agent" concept is a bit liberating and forces me to be more diligent about what I do in my sessions. I'm not terribly precise with my model usage but as mentioned before, Deepseek has been doing what folks claim GPT is capable of and GLM does what folks claim Claude is capable of. I spend way less - allowing me to feed my dog and fund leftist media.

I think I'll write more about my use in the coming weeks but right now I've been very focused on building out my Calibre + Goodreads replacement and when I give the progress update on that, I'll write more in detail about how I've been managing project development with it. It's been something akin to a local JIRA + wiki system, but with me effectively talking to myself.