A Month Off Code
Its been exactly a month since I landed in San Francisco.
Great place to be, lovely weather and an active buzz of keyboards clacking reverberates across all the cafés I've visited. And there is a lot to talk about this amazing city, music, food, parks. The vibe is nothing like what the media feeds down your throat. Yes it is expensive and there is the occasional whiff of piss around certain corners but there is a lot more to love than hate about it. I'm no writer, in fact this is one of 10 odd drafts of random writings ranging in technicality but writing is the best form of reflection and I wanted to document, for myself at least, the past month.
Now for a little background, towards the end of May, I had decided to take a month long abstinence from a new form of addiction that has hijacked all my dopamine receptors. Its consumption wasn't in the form of needles or pills, and its rather rampant in this city from what I've seen. Its been gripping me for the last 5-6 months ish and whats particularly scary is the mechanism of the euphoria it releases. Its doesn't really have a comedown, rather has a strong sense of progress linked to it.
Don't get me wrong I love it and continue to consume, however, in more mindful, mediated doses for specific occasions. What I'm referring to ofc is Coding agents, in my particular case, its in the form of Claude Code but other similar coding agent harnesses have the same effect.
Back to when I first landed in the middle of May, I was deep in building for one of my clients when I decided now that I'm here, let me see what the scene has to offer and sign up for a hackathon. Now on a foggy Tuesday evening in June, I get a luma notification telling me its tomorrow. No idea of what I want to build, 2 weeks into my no code break where I was trying to write more, put into words the one concept I thought I was an expert on in 6 months of inferring them. Suffering from the withdrawal symptoms of feeling useless having not written (or prompted rather) a line of code and heavy with an existential emptiness of purpose, I show up. No idea, no plan, just a coffee fueled 6 hours of building and a half written concept of false confidence that I knew agent harnesses enough to make my own.
Over the last 6 months I spent perfecting (to the best of my ability) my coding agent setup, handwriting custom skills and workflows for various tasks in my contracting work, trying various harnesses, trying out local models (well kinda, I need to spend more time trying Pi with Kimi/Qwen). This setup was still there and waiting so I got to prompting. I spent the last few months building for others, and yes I did build the odd productivity app here and reading list app there, but never did I truly sit down and think, what do I need. This was the opportunity to do so.
I wanted to build this but for people, humans. An Agent Harness not for your coding environment but for your life environment. The people, the tools, the interactions, every individual is heavily nuanced and I have been circling this concept for a while.
My idea was to build an agent personalized to the user, where initially it asks some questions to build an understanding of the human and their nature, what their likes, dislikes are etc. and then based on that concept create a few suggested actions that the agent could perform for the user like track a price on a pair of soccer shoes, summarize an article or newsletter, spin up their own skills and maintain its own memory. And I wanted it to work on the concept of always-on agent loops. But always-on price alert loops on shoes are expensive, a few tools are not enough and everyone needs a different set of curated skills that aren't obvious to the user themselves sometimes.
But in the build I realized these agent harnesses are like all the best technologies. They seem simple enough from the outside, you can visualize them, give them long elaborate analogies, breakdown their inner workings into a few core gears, but actually building them is easier said than done.
The problem was I had fluency in what I was using and understood its mechanisms, but lines of code and words on markdown files felt productive when I couldn't explain it to myself in the mirror. I realized this when I got called up to speak, being in the top 10, and despite putting on a good presentation, I was finding gaps in my own understanding of how prompts were called or messages streamed.
Recently both @kepano and @satyanadella had outlined the importance of grasping one's own comprehension prior to letting the parasite of false competence get absorbed and outsourced to these models. And first hand I experienced this via a 6 hour experiment rather than a 6 month long building sprint.
The long break away from building helped me appreciate and realize this. I'm still in my coding hiatus (outside of these hackathons) and writing has helped formalize some of what I knew to be true, but I look forward to exploring individual, personalized agents for each person's life or deciding to build something completely new. The economics are catching up where it's becoming viable to build. I moved here to build, but first I will spend more time understanding.