Two intense weeks, and a jig
Hey hey,
The last two weeks have been intense and beautiful. I’m super proud of the epic work my agents and I are delivering, and I’m working harder and longer than I have in a long while. Personally, I went to a gorgeous wedding in an iconic location, reached a new weekly PR on my running miles (27, ramping up slowly to avoid injury), and luxuriated with my visiting family.
Because of everything going on, I skipped the last two weeks. When I rebooted this newsletter, I promised myself I would try for a weekly cadence. I also promised I would not turn that cadence into a stick to beat myself with, or push out crappy writing just to check the box.
Instead, I took care of the things in front of me, made some real progress on work and home projects, and let the letter wait until I had something I actually wanted to share. That’s a more resonant note for the Thriving I seek to cultivate.
In motion
Proactive family finances
The Financial Quest is starting to feel less like an elaborate data cleanup project and more like a product my wife and I might actually use together.
The emerging goal is very simple: a productive, proactive household conversation about money. What happened? What needs attention? What should we do next? The system now has a home view, history, trip and tag exploration, planning, and a small attention queue - all designed to surface a few useful choices instead of turning us into amateur accountants.
There is still plenty to iterate on, and I’m excited about the direction. I need another handful of hours, and then comes the real first test - sharing it with my wife!
I’m very bullish that, developed properly, agents can help us cultivate financial freedom much more effectively than managing it manually. As long as I don’t burn all my cash on the tokens to track how I burned the cash, that is …
Agents raining from the cloud
This past week I went down a dubiously enjoyable rabbit hole on asynchronous cloud agents: persistent agents that live somewhere other than my personal machine, can work with different models, and leave traces and evals behind so the system can actually learn.
I often quote Mike Taylor’s Eight Levels of AI Adoption (opens in a new tab), and a few tipping points at work forced my hand to climb the mountain of Level 8. Definitely not trivial: I want agents with visible judgment, feedback loops, and gradually earned autonomy.
The current work is designing and building one very narrow use case end-to-end. Not a launch as such, but it feels like an important next layer underneath the more specific agents I have already been building.
Model + harness updates
Much to my own surprise, a few weeks back I got talked into trying ChatGPT. Even with Fable at hand, many of my colleagues and AI folks I respect and follow online were raving about Codex (as a harness) and ChatGPT Sol 5.6 (as a model).
I’ve been a monogamous Anthropic fanboy for ~18 months, and an ardent Warp + terminal touter for most of that time. “A purist!” I’d proclaim proudly, waving away native apps as somehow inferior to the cold, hard reality of the text-only interface.
Long story short, I’m converted. Codex as a harness is clearly superior for most of my professional and personal work, and Terra + Sol on High are both strong daily drivers. I find Medium or below makes weird, annoying errors or falls back into AI-ism slop.
Fable is still “smarter” in many regards, and I like the vibes at Anthropic much more, but it’s slow as a model and has its own annoying quirks. Plus, it’s ssssoooo spendy. I’ve also defaulted to the Claude app - even though I like it much less than Codex - because it has solved many of the rough edges I experience in terminal land.
When I have a super meaty project or something intellectually dense, I’ll have Fable draw up a plan, Sol review it, and iterate between the two until they converge. Then I’ll have Sol or Terra build it, Fable review and file issues, and Sol or Terra review and implement them. I iterate that way until the thing is built and stable. It’s a bit clunky, but I find the work product quite strong when I juggle the models.
It’s wild how quickly all this is changing. At work, I’m starting to fiddle with OpenRouter and am right on the precipice of turning on open-source Chinese models for less sensitive tasks. Again, folks I respect say they’ve gotten really good and, at a fraction of the cost, are a compelling alternative.
This is a big part of why I want to get my routine agent work into the cloud with traces and evals: once I have that infrastructure in place, I can more efficiently test different models to land on the dumbest, cheapest model (and harness) that’ll still do an excellent job. Heck, I can even imagine a world in the not-too-distant future where I have a dedicated LLM box at home to route my private or sensitive stuff (e.g., finances and medical). Building toward that Thriving agent-assistant future, one infra brick at a time.
Taking in
Designing With AI? Make a Jig. (opens in a new tab) by Jack Cheng gave me a phrase I have been turning over since I read it: make an AI jig.
As Cheng defines it, an AI jig is a simple fixture that helps you make the same kind of decision or action well, again and again. That feels like a much more interesting direction for AI interfaces than trying to recreate every old design tool in miniature. I was especially taken with the examples of direct manipulation controls - sliders, toggles, color pickers - attached to an AI-built interface, and with the idea that we are still learning our relationship to this new material.
I love the concept. It feels related to the systems I am building: not an AI that makes all the choices for you, but a thoughtfully shaped environment that helps you make better ones.
Code was our medium for thought (opens in a new tab) by Amelia Wattenberger is one of the most beautiful reading experiences I have seen on the web. It also gives language to something I have been feeling: code - like writing or a work deliverable - was never only output. It was the medium for exploring a problem, making decisions, and gradually turning fuzzy thinking into something clear.
The piece asks what happens if agents give us a custom whiteboard or playground before we even start with code. I love that. Thriving with AI means unlocking vast capabilities for both production and thought that simply weren’t possible before.
In a similar “new world of virtual storytelling” vein, one purely aesthetic recommendation: Laura Entis’s Inside OpenAI’s Race to Reinvent Software Development for the Agent Era (opens in a new tab) at Every. The thinking is aligned with what we’re building at Coval, and more to the point here, it is beautifully designed. The dots are a lovely through line, beginning with an agent’s late-night Slack message about a failed export. Worth a read just to enjoy how a digital story can be made.
Next week: a smaller, more playful AI quest. It is almost ready, and it’ll be fun to show you.
Until then,
~h