Notes for August 29–September 5

As many people have remarked, the moat between ideas and technical execution is narrower than ever, and that has a number of consequences–both when it comes to investment (in this case, personal, although I am collecting war stories…) and focus. And I think I need to be more selective about what gets my evenings, since I am back at work and already spending too much of my free time building things on a computer.

Why? Because I am right back where I was before my break, spending too little time doing anything but sitting at a computer, or using a computer to do things that only work on a computer.

I keep having ideas, and even though I finish my workdays achy and drained, I inevitably get caught up in the excitement of being able to build stuff quickly. And since my free Codex subscription is due to end this month, I’m trying to make the most of it…

I need to get back to electronics, hardware, and reading (maybe even music) as a hobby, and stop spending so much time glued to a screen–which is something I have been trying to balance against my writing backlog, and part of the reason the notes this week are relatively sparse.

But there are some hints of progress in finding a better balance between screen time and other activities:

I finally hit my daily step goal for several days in a row, which has been a small but satisfying victory and has eased some of the strain from sitting at my desk for long periods (albeit not completely).

The trick? Doing the grocery shopping myself. You end up walking more than you would if you relied on delivery, and it adds a bit of variety to your daily routine. Not to mention milk carton lifting, which is a surprisingly effective upper-body workout.

I spent some time (re)setting up test machines and going through my review backlog, trying to catch up on notes and tasks that had accumulated while I was focused on other projects–even though I have little to show for it in terms of output, I managed to at least organize most of my notes, some of which date back several months now…

As an offshoot of the work I did with micro-VMs, I’ve been poking at two related problems for a while: getting agents to preserve their state, and tackling the holy grail of freeing computing environments from being tied to a particular machine.

legion takes that idea and gives it a twist. Rather than packing everything into one portable process, it is a self-hosted runtime for durable agents backed by a Raft cluster, with WASM and Bun functions (making those available across the whole cluster is still work in progress):

• every agent turn is event-sourced
• nodes form a Raft cluster
• code is deployed as content-addressed WASM modules or Bun bundles
• it’s all organized as a 9P namespace

The namespace provides a uniform way to manage the whole thing, and was, I confess, a whimsical choice. But my intention is that crashes, restarts and individual machines going away should not make an agent stop or lose track of ongoing work, and so far this seems like a moderately sane way to do it.

This is still at an early design stage and might end up going nowhere, but there are already enough runnable examples to exercise most of the architecture:

On a more pragmatic note, I have a “me” problem with Azure Virtual Desktop: meetings happen inside it, but sometimes I need to share something that is on my local desktop, which the remote machine obviously cannot see.

My poor man’s fix is provisionally called ShareCam: select a region locally and pipe it into AVD as a camera. Crude, but it works, and once I clean it up I will put it up on GitHub someplace:

And yes, this shows up instead of my camera view, but that is perfectly acceptable for my use case. The only real challenge is that screen sharing typically has a dedicated pane in Teams and Zoom, whereas my solution just replaces the camera feed and people have to arrange it themselves accordingly.

This was also based on rcarmo/swift-app-template, which I’ve been tweaking to build all my new little Mac tools. It took me all of… 25 minutes to get it working end-to-end, including a second pass for individual window selection and capture that still needs some UX tweaks. gpt-6-astra did the core implementation in less than 15 minutes–having the template ready certainly helped.

And this, I think, is what I should be doing more of: investing my personal development time in satisfying life hacks and little tools I will actually use instead of trying to boil small oceans–those I can always keep doing at work…