State of AI Assisted Workflows July 2026 | Handmade Oasis
State of AI Assisted Workflows July 2026
27 Jul, 2026
• 3260 words • 16 min read
Three months since the<br>last update, give or take a day.<br>In that post I landed on the idea that the layer around the model is where the<br>leverage is, the harness, the context, the review loops, all the boring pieces.<br>I still think that’s true. What changed this quarter is which problem I point<br>all of that at.
Because what I have actually been optimizing for is not speed of generation<br>anymore, it’s keeping my mental model current at a rate of change that used to<br>be impossible. That’s the whole post really. Everything below is either<br>something that helps with that, or something I stopped doing because it quietly<br>worked against it.
I did not arrive at this by thinking it through. I got there by trying the<br>opposite first, taking myself out of the loop almost entirely and letting the<br>thing run, then noticing what that did to me over a couple of months. More on<br>that further down.
I’m at somewhere around 30-35B tokens total now. That’s barely up from where I<br>was in June and I think that flatness is the most interesting number in this<br>post. I use these tools more than ever but I have also gotten way better at not<br>wasting tokens. I am very deliberate about what goes into my system prompts,<br>skills and what I expose to the agent through my harness.
What I’m currently doing
Harness and models
Still Pi, and I don’t see that changing. Nothing I have tried comes close in<br>terms of how much of it I can bend to how I actually work. My defaults right<br>now are 5.6 with thinking on high, and I reach for Opus 4.8/5 quite a bit, the<br>frontend gap I mentioned in April has mostly closed but I still find the<br>Anthropic models nicer to think out loud with.
I still keep Codex and Claude Code installed and use them regularly, but at<br>this point they are observation posts more than tools. I want to see what the<br>labs think a harness should look like, because they are the ones RL’ing the<br>models against their own scaffolding, so when they add something it’s worth<br>asking why. Then I go see if I want that in Pi, and usually I can have it in an<br>afternoon.
I do use the ChatGPT desktop app quite a bit now, it’s got really nice UX for<br>things that are not code related and that I don’t really care that much about.<br>It is admittedly still way too bloated and does include a bit too much in its<br>context as system prompt and tool schemas and skill descriptions etc but all<br>things considered it’s still useful.
The thing I use most in Pi that I did not really use in April is the tree<br>structure, properly this time. With auto-trees and a /prime command to set up<br>the root context, the branching stopped being a feature I knew about and became<br>the actual shape of how I work. Go down a path, hit a dead end, go back up,<br>take the learnings with me, branch again. Sessions are cheap, being wrong is<br>cheap, what would be expensive is carrying a polluted context forward and<br>having it interfere with the task at hand or just fill the context window with<br>unrelated tokens.
Herdr instead of tmux
I have used tmux for years and I have now fully replaced it with Herdr. It’s<br>clearly built by someone who works the way we work now in this new era of<br>agentic coding. It can do all the things that tmux does in a very nice way but<br>the new capability that makes it very valuable for me at least is that agents<br>can drive it, which sounds small but changes what you can automate. As an<br>example one of my favorite things to do is have the main session agent spin up<br>an explore subagent in a new Herdr pane, that way I can see exactly what it’s<br>doing and also even steer it since it becomes like a normal session that<br>happens to talk to the main agent through Herdr, it’s phenomenal really.
Herdr also adapts to a phone viewport really gracefully when I’m connected over<br>ssh to my machines. That last part matters more than I expected, which brings<br>me to the next thing.
Working from my phone
I set up Termius on my phone and a remote-access script on my laptop, which<br>brings up a user-space sshd, mosh and keeps the machine awake, all over<br>Tailscale. So now I can be out somewhere and drop into my actual machine, with<br>my actual environment, my actual keys, my actual repos. Not a watered down<br>mobile app pretending to be a dev environment, the real thing.
I did not expect to like this as much as I do. It’s not for writing code, it’s<br>for the moments where you’re away from the desk and you want to kick something<br>off, check on a run, or capture an idea before it evaporates.
I do enjoy dictating over SSH mostly, and for getting ideas out of my head fast<br>when I don’t want to type on a clunky phone interface. Note that this is all<br>non-coding use. In January I wrote that friction between thought and<br>instruction is a feature and not a bug and I...