One observation
The rate of change in the agentic world is astounding. Every week there’s new, compelling desktop apps with shiny websites promising to run every aspect of your business. And every day, the model providers are launching harness updates and behind the scenes model changes. Anthropic had 20 major releases of Claude in August. The tool I’m using today is different than what I started with in June. I first drafted this note 9 days ago and I’ve got a few more tools to review.
There’s an understated risk of being left behind in tooling. The models are only one element of the workflow and lab. Apps wrap the models and put it to work and how the model is used makes significant difference for the results. Models in and of themselves are insufficient for creating software.
The harness
The first thing that I did when I started this journey was download a harness. In my case Anthropic had caught my attention so I installed Claude first. It is a cli so I was very comfortable with adding it to my existing iTerm (and now Ghostty) workspace.
I’ve also tried ZCode (desktop only), Antigravity (cli and desktop), Orca, Zed, OpenCode, OpenClaw, and probably a few others that slip my mind. Just about everything except Hermes and Codex I suppose. Some don’t last more than a launch or two. I have little tolerance for bugs or flaws and I expect to be running very quickly.
The first thing I look for in a harness is if it’s configurable to use any model provider and any authorization. I do this because I’m convinced that using multiple models from multiple providers is efficient and economical. (More on that later.) I’ve also found that subscription billing is more economical for the way I work. I suspect that’s largely the case for engineers - it seems like subscriptions are a deal if you regularly max out your quota. I also want the harness to use the provider directly - I gave OpenRouter and claude-code-router a test but I found it flakey and I wasn’t interested in spending a lot of time wrestling with a new point of failure in my workflow. I’ve developed a very strong preference for Claude Code cli now and it’s progress. It’s usually where I start my work and I found a nice configuration hack that allows it to use other providers like Z.ai which has implemented the Anthropic call interface.
I also use ZCode from Z.ai as a GUI harness for GLM-5.3. This is another product that has progressed rapidly into a very nice interface. One nice perk - Z.ai has been giving away a lot of access to their Flash model on the weekends. I was able to complete two very large rounds of refactors on PPPS using their bonus time. I presume I’m providing training data but I’m not concerned about that on my projects (yet).
I tried Zed after reading a few good reviews about it but it was lack luster.
I was intrigued by OpenClaw, in part because I’m familiar with Peter Steinberg and his work on iOS apps back in the old days and he’s a very capable architect and creator. I haven’t used it much though as I didn’t see the advantage of it over Claude Code itself. He’s since joined Anthropic, presumably to work on CoWork and Claude. Funny story - as I onboarded I gave it my name, a bit about myself, and told it to act a bit like the computer from Star Trek, be concise and brief, without flourish but occassionally be witty. After on-boarding it recapped and for my Name it had “John Fricard”. touché witty claw, touché.
The difference between large team products and startup/small team/solo projects is noticable in two ways: rate of change and accuracy. With some of the hyper-promoted apps like Orca and Zed there was a significant gap between the docs and the app. Documents getting stale is an age old problem in software, but come on - we’ve got sythetic workers available now! Updates are one context away.
The rate of change coming from large teams is astounding, as you’d expect. A test driven team, with a clear understanding of how to really engineer software can generate a lot of stable code very very fast. And if you’re using the tools you are building to build the tools you are using then your own context cycle is going to be very tight. Claude Code is open source so you can see how it’s going.
This note will be ongoing as more things catch my eye and I give it all a try.