How I 'write software' as of Oct 2026 - Part 1

Where do I begin?

Everyone has been writing these awesome "here's my coding setup" blog posts, and I've wanted to get in on the action for almost a year now. Problem is, as I'm sure many of you can relate, my setup is changing almost as fast as everything else in this industry.

For quite a while in late 2024 and throughout most of 2025, I was happy just coding 'the old fashioned way' in VS code or Zed with a bit of extra juice from cursor, and later, Zed's AI support and Antigravity. Then, like many others, I started leaning more heavily on claude code, opencode, gemini cli, codex, etc. For about two months in late 2025/early 2026, I felt like a wizard with my own vibe-coded version of claude's remote control feature before that became a thing; mine worked through tmux and shared a lot in common with Orca. I called it "Vibe All Over" and I thought it was just the coolest thing. I could control claude code from my phone! ANYWHERE!

But then claude code remote-control did come out and for the next few months of early 2026 and into summer I was a big claude code remote control user - until I discovered Paseo, codex got really nice with the 5.5 and then 5.6 families and things really kicked into a higher gear.

But to understand these changes and where I'm really at now, you first need to understand the fact that computers aren't just a means to an end for me, they've also very much been a hobby as well, ever since I first got to play some DOS games like Digger and Flightmare on my Dad's Columbia MPC 1600 back in the 80's.

As fun as it'd be to indulge in some computing nostalgia, this blog post isn't going to focus on my computer history lest it become far too long. So let's just jump to the part where I'm well into my 40s and suddenly there's a new kind of AI in the form of these crazy powerful coding agents like claude code and I finally have a good reason to keep that little $150 tiny PC running more often doing something.

My fleet, for now

Serve    core i3 and 14TB of ZFS in a shoebox: Tracker, Forgejo, UsageHub, Tasker, k0s coordinator
Forge    9800x, 32gb ddr5, 2x 20GB 7900XTs - local private inference - ~120 tok/s qwen models
Window   96gb ram, 7800x3d Windows desktop + several WSL2 distros; main workstation
Bench    strix halo 64gb tiny but mighty "laptop" (technically a tablet)
Lab      small but capable dev box - notably has 16 cores, nvme + a 2.5" sata ssd
Legion   beefy legion gaming laptop with 16 cores and 32gb ram
WMBP     M5 Pro + 48gb
AOC      tiny intel atom box, good for k0s coordination  
hpthin   hp thin client workhorse - k0s coordinator, low power, 32gb ram, 2tb nvme ssd, vpn
phone1   paseo app, (now crusteo) works with zerotier
phone2   my coworkers have been known to make fun of me for rockin' two phones

I can't say if it's because I had the computers I felt compelled to use them for things, or because I started seeing more and more useful work come out of coding agents that the ideas for all the things I wanted to build suddenly seemed possible if only I had enough hardware. Most likely it was neither; it was probably my default drive for frugality seeing a Claude weekly quota potentially going un-used and thinking "better give claude another task, however silly.. think of the unspent tokens!"

Context Organization

So how do I organize these computers and my coding agents around work? Well, mostly, everything revolves around tracker, but saying that doesn't really give you enough info to understand how things actually work.

Thanks to the fact that I have a few long-running projects like zpq, I've learned that you really need more than just a codebase to effectively utilize more than a couple coding agents (just like teams of engineers tend to need wikis and ticket tracking systems, I suppose..)

The thing is, I didn't want agents to have to waste a lot of time doing more work than what's necessary interacting with those tools, so I just have them write notes to themselves (and me occasionally), and aside from telling them where tracker is they seem to just 'get it' almost scarily well.

If you're wondering how I prompt my agents to use these tools, I wish I could say it's just some skill you can copy from me or something, but the truth is, I just talk to my agents like they're fellow teammates. When starting one for the first time in a new project, I describe my goals, foundations of the design I know about, chat about various tradeoffs if applicable, and build up to a comprehensive program design first with a high-tier model.

Once that's in a good place, I'll typically refer to that agent as the "chief architect" or similar and spawn a tpm of sorts next, and tell them that they'll be my primary interface into the project. I'll introduce them to the repo and the chief architect, then the tracker folio where the work is being documented for me and all agents. From there it really depends how big the project is but it's at this point now where my new paseo-compatible front-end really comes into play:

Wrangling the Swarm

Crusteo is finally making a dent in fixing the biggest problem I've felt like I've really had since all this stuff started getting so capable around early 2026 - that I WANT to have a lot of stuff in flight, and I want to be able to keep making little bits of progress on each item, but that if I don't organize all these threads efficiently in some organizational system, it gets too easy to lose track of what's doing what. My friend Jos said it better just a few hours before I wrote this post:

Knowing what you were doing is half the battle, and seeing all the related agents together in the same lane has almost completely solved this problem for me. And then there's UsageHub, which is of course integrated right into crusteo as well, so I know whether I should kick off that extra research now or not:

Don't get me wrong, vanilla paseo might be fine for you and I highly recommend just trying it out first, but for me, a huge scrollable vertical list and just one agent panel open at a time was not helping as much anymore. In fact, I was finding myself just opening 6 or so instances of paseo at once to keep each window up somewhere on my monitors to watch it along with the others until I remembered that we live in the future now and I can just make things work the way I want them to instead of putting up with something less ergonomic for me.

So now, along with a desktop view that looks more like a high-frequency trading application letting me quickly pan left and right through all my different projects, I can see which agents are working very easily and drop right into any of them to pause, steer, or just observe.

And best of all, as of today, I have almost 80% of the same experience from my phone, with an interface I've been honing with three levels - a zoomed out map of all agents placed into their according project, a filterable list view, and then an optimized one-agent view that lets me interact even better than claude code's own remote system:

I'm still polishing these pieces, but soon I'll have this on the crusteo repo as well, where I'm making not only a native drop-in replacement for the daemon, but a brand new middle-tier hub which you can run wherever you want to centralize client connections to your daemons so you can stream it out efficiently to the web client, which you're seeing screenshots of here.

There are some more parts to discuss, like my local inference server setup, a slightly-undercooked tracker extension I'm really excited about, and how my fleet records every turn that'll have to wait until part 2.. but as always, I'd love to hear your thoughts, comments, and questions!