I’ve been all over the place lately, coming up with ideas, trying them out, throwing away more code than I actually kept, etc. And this whole firehose of thoughts has uncovered another issue. We’re in this weird transitional phase where the old deterministic way of doing things and the new “ask an LLM to figure it out” way are both valid approaches to the same problem. And figuring out when to use which one is its own challenge. ...
In using the tools I’m limited with at my day job (github copilot, yuck), I’ve had to find some workarounds to the built in difficulties. Microsoft is doing a lot in that tool to really limit its costs, which, sure. But it does make this tool much less powerful than the unfettered tools and models (and I’ll point out, every time a developer has a bad experience with Agentic AI tools, you risk losing a customer for life). They are doing things like artificially limiting the context windows down to 100k tokens, adding stuff to the system prompt to make sure it doesn’t run for too long, etc. They’re obviously also using the “dumber” models and just calling them by their standard name, for instance there is no way in hell that the GPT-5 model in GHC is anything but the medium version or worse, because that thing compared to what I use on my own time is significantly nerfed. ...
One aspect of this new paradigm I’ve been really enjoying lately is rethinking all of these things we’ve just taken for granted for years. Like the fact that all of our tools are designed for humans. This usually will result in inefficiencies, such as wasted tokens going back to the model, or giving it more to sift through. But overall, I think we can do better than rg for search or curl to read a website. This led me to start building some new tools, to try this out. ...
Every time I start a new project, I tell myself, “I’ll document this properly from the beginning,” and then three months later I’m staring at a README with a single line that says “TODO: Write documentation.”. Eventually I got tired of that loop and started tinkering with a script to help me out. The tinkering snowballed into Wiki Bot, a little assistant that reads a repository and spits out GitHub wiki pages. Just started working on it, but it’s been surprisingly handy. ...
It’s not all doom and gloom. In fact, I’m pretty excited about this change. It comes off weird sometimes because I, and most developers, I think, have a lot of thoughts about this new world. It’s a lot to process. I’ve spent the last, I dunno, four months chewing on what this means for my career, the industry, everything. Eventually I decided to start writing about it, and that’s how we got here. ...
As I write this, I’m watching a team of Codex exec runs working in tandem and in sequence to build some software I had an idea for. They start with a backlog-mining phase, plan out their own work, create GitHub issues for each piece of work, then another workflow kicks off that starts with a planning phase, after which the coding agent implements the change, the testing agent creates tests, the documentation update agent updates the references, and an agent handles the PR. A whole agentic software development system that churns through the idea, non-deterministic token output at a time, until we get to something that is not only usable but well-architected and maintainable. ...
I built something that still feels a bit surreal to watch in action: an orchestrator that chains together AI agents to handle the entire software development lifecycle. From planning to deployment, it’s agents all the way down. Honestly, it’s not even anything too impressive or hard to build at all. Just some simple python to handle a workflow system, and to wrap either the claude cli tool or codex cli. The really crazy part is that once you have the idea, it takes an afternoon to get it up and running. ...
So I’m in my comfy chair with a new MacBook Air, doing what every normal person does on a Friday night: browsing Hugging Face. I got sucked into the audio processing rabbit hole and realized that, despite years tinkering with RNNs, regression, and the usual computer-vision suspects, I had never tried to make a model understand sound. That felt like a gap, and I hate gaps. Whenever I pick up a new domain, I give myself a project that seems “reasonable”—something I could finish in a week if life doesn’t intervene. Fresh off editing a YouTube video, I fixated on how much time I spend deleting filler speech. If you’ve never lived inside a video timeline, disfluencies are the ums, uhs, restarts, and other verbal trash we shovel out so the audio flows. For me, it’s mostly stripping the words “umm” and “so” until my waveform looks like a comb. Naturally, I thought, “Hey, maybe I can teach a neural net to spot the ums for me.” That sounded perfectly doable. ...