Notes from the agent coding (not quite) vanguard
Everything is weirder and more capable. Models are way better at many things, unattended coding in particular. But, to everyone’s dismay, that causes them to write in a tedious style of English.
There’s plenty of room for humanity in between the tokens.
1. Models & harnesses
Frontier lab models, despite their quirks, feel like they will graduate to boring technology soon. When the servers are healthy, you can rely on Anthropic and OpenAI’s models to get work done. Cursor’s in-house models (probably other labs too, I haven’t tried Kimi, GLM, etc.) are solid enough to pinch-hit for frontier models when the inference servers are buggy.
Harnesses, when they’re relatively bug-free, are basically just as reliable. You can use the frontier labs (I’m still on Claude Code, mostly) to get most sorts of stuff done. You could substitute Cursor, Pi, or most of the others to personal taste. And, personal taste is increasingly a thing across the harnesses. Claude Code is very much for the terminal people, Cursor is increasingly for the holistic product people, the others are for the customization enthusiasts.
Occasionally one model will make a noticeable leap over the others in capability, speed, or “workhorse-ness”. Occasionally, a model gains a unique capability that is a big advantage until it’s cloned into the others. I find those momentary blips don’t matter enough to switch between models and harnesses that often.
(Despite that, I’m a novelty junky and I still oscillate between Claude Code, Cursor, and a couple of other approaches more often than I would like to quantify.)
Optimistic prediction: there will be a lull in frontier model capability race in the next 6–12 months. It will either correspond with IPO quiet periods or increasingly bad optics about the development of increasingly autonomous models. I’m optimistic about this because I think we could also use a breather in this domain.
2. Skills & prompting
Models and agents are good enough, in the late summer of 2026, to focus on customizing and leveraging skills instead of the model of the month. Bonus, skills are portable across agents, and models, sort of.
For most folks, those living in the middle of the bell-curve (definitely not the Amp folks, who seem to exist in the future), I think this means the action is now about telling the agent what the task is and how to figure out if it’s done or did a good job. Prompting and verification.
I’ve had a lot of luck lately with simplifying my instructions and moving more of the details into (Markdown) documents. In general, a lot of what I had in my CLAUDE.md ten months ago feels like over-steering now.
That said, moving the quasi-deterministic “verbs” of your interactions with coding agents to skills still makes a lot of sense. In particular, to socialize usage patterns with teammates or codify how specific tools, e.g., Xcode or CSS, are used.
Optimistic prediction: most currently-popular skills will be trained into the models. Instead of /grill-me about the new login screen or /fix the bug where the button disappears evaporate into garden-variety English.
3. The human factor
A couple of weekends ago, I was tinkering with doing Claude code via iPad and remote control. But I remembered the smarter move is to avoid tokens on evenings, and probably weekends too.😇
Lately, I’m finding that the drive for software factories and sprawling specs feels too much like the boring kind of software development work. Nudging and steering an agent, asking questions and getting answers in the form of working code, that’s more fun. It’s not as surprising and energizing and sometimes-frustrating as human collaboration. But it doesn’t feel like doing rote homework problem sets, which must count for something!
Optimistic prediction: the models whose output is most concise and least cliché could prevail over more capable but inscrutable ones.
4. Even more predictions
In the coming months:
- More hobbyist/self-employed/solo developers will utilize >1 agent harness monthly subscription. For bouncing back and forth between whichever model is in the out-performing at the time. Or if one prefers “the vibe” of one model as a collaborative partner while using the other as a coding agent. Sometimes, to get extra token usage at $40/month instead of going all the way to $100/month. More reasons to carry subscriptions to two or more models will probably come along as use-cases evolve.
- As more developers carry access to more models, skills and workflows will gain traction that use models in an “adversarial” mode. That is, model A reviews model B’s plan, model B review model A’s code, etc. Or, model A/B/C all try to find bugs/security issues in existing code in a competition. I’m not sure how this will pan out, but I’ve heard of people having good luck with this currently.
- One of the frontier model labs could go public. We get a better idea about the economics of their operation, what token subsidies actually look like in numbers, the actual economics of training and releasing a new model, and what the labs think their long-term play is currently. (Pessimistic: their filings and road-show are just “come with me, it will be good, but we’re doing it my way” like Facebook, and we don’t learn anything.)
- That developer with Pi kitted out just like they like it is the new co-worker who insists on vim (not NeoVim) or emacs with their dot files just like they like it. Same as it ever was.
It feels boring to write about agent coding so often. I try to keep it spaced out, even if it’s with modest vacation photos and ramblings. But, so much is changing and often interesting that it feels like missing out on the moment to develop one’s thoughts, in writing and publicly.
So, I regret to inform you, the discourse on AI will continue until the situation becomes less dynamic, fascinating, and rewarding to develop what I think of it by writing publicly.
Interestingly
Three modalities of creativity I am keen to explore: dirt, screenshots, monk-like.
1. Dirt, metaphorically
Take something worthless (e.g., literal dirt, metaphorically non-working code), put great effort into it, get something nice in the end:
Sometimes things don’t go as planned and the product that comes out the other end is really not what I wanted or needed. At that point, the right thing to do is usually to start over from the original specs (and possibly the wrong code) and restart the spec and design process. Then implement again from scratch. There are absolutely projects that I’ve run through this process five or six times as I figured out what I actually wanted or the right way to explain what I was going for.
Dorodango is, essentially, the process of polishing a ball of dirt into a beautiful, high-gloss sphere. The result is genuinely amazing.
— Jesse Vincent, Dorodango
2. Screenshots
Knowledge workers are all taken with our text files, spreadsheets, or design files. We invent ideologies and workflows, have very specific opinions about how to do this thing, know all the keyboard shortcuts, organize them just so, or not at all!
And then, there are eccentrics who have these giant collections of screenshots, thousands of them. Of order screens and login flows and memorable conversations and receipts and profound quotes on the web and sports scores and seventeen iterations of one document/screen/logo.
During the opening of the conference, Omar honed in on the subversive nature of the screenshot. In popular computing, it circumvents the app siloes that define our contemporary digital ecosystems. A screenshot doesn’t need a log in, bypasses DRMs, and is interoperable between practically every single computational device. Even in the “high-culture” of computing, where text is dominant, screenshots prove subversive. The screenshot is unstructured information which must be parsed to offer the tidy data best suited to computer hacking. They are seen as verbose and unwieldly, despite universal adoption. The Tao of Unix is text files, not bitmaps.
— Cristóbal Sciutto Rodríguez
3. Monk-like devotion to the craft
Deep, monk-like devotion to the craft of software development and its output, the application. I love an in-depth, illustrated essay on decisions that go into a carefully considered application.
The initial vision for Paper was simple — build something that has the core tricks of iA Writer, but in a package that feels even more elegant and minimal
…Have people noticed the effort? Most — probably not… but some have.
I don’t have a big connection to make here! Deep attention to detail, knowing when to polish and when to start over, trying to make different media work – that’s it. Make interesting things, interestingly.
I’m on book six of Dune. But, I’m just now finding out the story runs eight books. (Have Frank/Brian Herbert not heard of trilogies?)
Finding focus when working with coding agents, flying too close to the multitasking sun. But, it turns out concise output and taking intentional time away from tokens is the key.
Driving more confident in autocross, trying to use the pedals, both brake and gas, more. But, I’m still slow.
It’s all dumb. But, we learn things.
Watched: Spider-Man: Brand New Day 🎥 Properly rated at ~71% positive. It’s a fun movie. Definitely better than the second Spider-Man, probably better than the third, not as good as the first. Only 2.5 hours, but it felt way longer. There’s a bunch of plot movement to get through.
Holland and Zendaya still have great chemistry. They pulled in a few minor characters from three of our favorite long-running dramas: Michael Mando from Better Call Saul, Liza Colón-Zayas from The Bear, Tramell Tillman from Severance. Jon Bernthal is surprisingly likable as The Punisher, which I don’t even know how to feel about.
Is Marvel back? Possibly! Spider-Man is definitely the best world they have going. But the supporting cast of shows and characters appearing in the upcoming films may make the post-Endgame dip feel mildly redeemed if they can land Doomsday. They’re pulling in the X-Men universe too, ever so gradually. But not blowing it — yet.
Patrick Dubroy, Fast is better than slow:
Think about it — if you’re fast, you get data more quickly. That helps you make better decisions, sooner. It also means you learn faster, and over longer periods it means you learn more. Being fast also means you can try out multiple approaches to a problem and pick the best one.
Fast is more learning is better execution.
Don’t worry about looking dumb. You probably already know that you should share your work early and often. But it’s uncomfortable, so it’s easy to put it off while telling yourself a story like “I have a high bar for quality.”
You’ll get results much faster if you learn to push through that discomfort.
If I’ve learned anything from working with coding agents on a team, it’s to worry less about that one code review where someone points out a howler of an issue with my code. Whether a coding agent generated that code or I typed it in by hand, acting on my teammate’s feedback is an opportunity to show that I care about quality, I listen, and I follow through.
That’s what (still) makes great coders. And, fast fingers.
Frozen 2 should be rated R (Interconnected):
Frozen 2 in which the entire city is almost destroyed by a tidal wave… It is SO LAZY.
While we’re piling on lazy writing here: LOST, and most American television, is particularly lazy about generating consequence and tension with lines spoken with characters holding guns at each other. Do we need the cheap escalation of a gun to know that the character really wants to coerce the other to action? To paraphrase Laurence Olivier, did the writers consider letting the actors act?
(Granted, firearms warp most parts of American culture, but this one is difficult to unsee once you recognize it.)
If the clankers have you down, I recommend: read your favorites, write like that, discover new favorites, write and edit like that, rinse and repeat. It would seem, at this moment, only a human can produce liveliness by chaining words into language and narrative.
Writing is how I build. That’s why I write.
Working with agents, summer of 2026: you have an always-on, amazingly deep collaborator. The coding ones can work themselves out of most problems, if you are willing to pay for the effort and thinking costs.
But, wow, you are absolutely right! They are boring as nails and write like an entrance exam question.
It is a real relief to spend a weekend, touching grass or concrete, away from the discourse around maximizing, minimizing, or wisely leveraging them.
The frontier labs regularly run promotions like “hey there’s a long weekend or holiday soon! Tokens or this new model is half-off until the next business day back.” I don’t like the nudge to “always be prompting”, working with the models at all time. To trade a long weekend for better token rates smells even worse.
The new sabbath was an offline weekend or vacation. The new time offline is time away from tokens.
So what had happened was, my daily reading sort of tailed off. Distractions and all that. Then my writing dropped off. And then, the web hit me with three great ideas hit me right in a row. The web provides energy for reading, the reading transforms into energy for writing. Suddenly, here we are, posting the links again. The web is good.
Mandy Brown, Umyazu:
Nor do we read when we slip through the stream or flick through the feed. Reading is an awakening of attention, not a deadening of it. We read to come alive to ourselves, not to forget who we are or what we are doing, or what is being done to us without our consent. We read to encounter the world, to connect what we know to what we do not know yet, knowing all the while that such understanding is always temporary, lovely precisely because it is transient. The suspension of disbelief that a reader brings to a text is an openness to becoming someone new, to shedding old selves and wriggling into new ones. It is an invitation to change.
Austin Kleon, Problems of output are problems of input
When I stall out, it’s time to start taking things in again: read more, re-read, watch movies, listen to music, go to art museums, travel, take people to lunch, etc. Just being open and alert and on the lookout for That Thing that will get me going again. Getting out the jumper cables and hunting down a battery.
First installment of this could-be-a-series of monthly reading notes. I thought I haven’t been reading much lately, and there goes my self-awareness.
You (almost) can’t do too much reading. Lovely thing, that.
Grass touched, wind felt. Mountains driven, autocross done in ski resort parking lots. Bend, OR and its surrounding areas.
You ain’t got time for all this – the latest newsletter edition from yours truly. Some of my recent writing, remixed. Some all new links. If you’re seeing this on your computer screen, you might like it!
🎧 Human Touch is the Phil Collins-est of Bruce Springsteen albums. From one Oscar winner of the 1990s to another.
📚 Finished: In a Narrow Grave by Larry McMurtry. If you’ve ever lived in Texas, even more so if you’re from Texas, this is a great read. It’s not quite Texas as I remember it, as the author was writing a decade or two before my time. But it captures the spirit and backstory of 20th century Texas. Contains great explainers on the rivalries between Austin, Dallas, and Houston. And, that Texas is at least two different things: east Texas, west Texas, and probably south Texas.
The nostalgia was strong as I read this. But, the headlines bear many reminders that the religious right and culture warriors have thoroughly captured and corrupted the state. I hope those who’ve stuck around can recover it, someday.
The local maxima we need to reach the next global maximum
Half-way through 2026, I’m finally getting around to posting some predictions for the year. Really, ideas about how software development can fix the house to deal with coding agents generating lots of code:
- Design with pacing, stable/unstable layers – ala Shearing layers
- Correctness by design/assertion – ala “Making Impossible States Impossible”
- Formal verification (finally) – figure out if we can do better than compilers and unit tests to know if what we asked for resembles what we received
- Even more deterministic static analysis – absent a leap forward in the practicality and ergonomics of formal verification, can we manage the flood plain of incoming code with beaver dams of linters, style checkers, and good old-fashioned
grep? - Improve our observability stance – how can we see the emergent properties of the running system, let alone iterate on them? And, can we do so without spending all our lunch money on observability services and stacks?
Notably absent from this list: guardrails, specs, swarms, optimization, maximalism, orchestration. These aren’t even local maxima, they’re just shiny objects or shovels someone is trying to sell us along the way.
Exploring the Rich History of Funk and Soul Music: A Journey Through Rhythm and Emotion:
Funk and soul music aren’t just sounds; they’re vibrant threads woven into the fabric of musical history. Emerging from the rich cultural tapestry of African American communities in the 20th century, these genres have become a cornerstone of musical innovation. They blend gospel’s fervor, jazz’s complexity, rhythm and blues’ groove, and rock’s edge to create a sound that moves both the heart and the feet. Here, we’ll dive into their roots, evolution, and the enduring influence they wield, keeping audiences captivated across the globe.
I couldn’t have introduced the funk better myself.
Related: Maggot Brain is highly rated but still underrated. That album could stand against any of Stevie Wonder’s “classic period” albums.
🛬Back from Spain. Vermouth de grifo, Hieronymus Bosch and Matisse, old buildings, and public green spaces are pretty great.
📚Still at it with Power Broker and the first six Dune novels. Larry McMurtry essays are making me a little nostalgic for the idea of Texas (but not contemporary Texas itself). Contemplating adding the giant book of all the Montaigne essays to this list.😎
🎥 Hoppers was fun and feels like possibly a return to form for Pixar. Mandolorian and Grogu isn’t the best or worst Star Wars, but it is fun and hits all the expectations.
📺 Rooster is great if you’ve ever fancied becoming a sharp-witted author and visiting professor in New England. Widow’s Bay is great if ever you’ve wondered what being the mayor of an obscure but definitely cursed island in the northeast might look like.
⭐I should call out Hacks here. That is how you end a series, a comedy in particular. I will now enthusiastically recommend this show to everyone as The Ideal Comedy television series. If you haven’t, go watch it. Five seasons, no skips or lulls.
Fits on a floppy, great idea:
Software has lost its way. Apps that once shipped on a single floppy disk now demand gigabytes of your storage, minutes of your time, and far too much of your patience. We accepted this gradual bloat, but that’s not progress.
Software should be as small as it can be. Not as a gimmick, but as a discipline. The floppy disk is the measuring stick: 1.44 MB. If the software that ran entire businesses could fit in that space, then a modern, focused, single-purpose tool certainly can.
Yep, these criteria make the good stuff:
Launches instantly. Faster startup, nothing unnecessary to load.
Does one thing well. Focused features, fewer bugs, software that lasts.
But, I disagree on some details:
Native only. No dependency bloat, every line of code earns its place.
I usually don’t feel like “native” is crucial these days. Whatever your definition of native is, it’s not a prerequisite for building great software. Even for desktop/mobile platform development using the vendor-intended language and frameworks, the downsides are…a lot, lately.
You can create software that is fast and focused with any language or stack if you’re careful. So, I might replace this one with something about attention to details and care for the craft.
Runs on older systems. Older devices deserve love too.
I love it when this happens. It’s easier to make it work outside native environments, too!
I think the applications are less important than the data. I’d change this one to something about the ability to store one’s own data on a floppy-sized local file instead of in an opaque cloud or a row in a SaaS database somewhere.
Maybe the big idea is that the application and data should fit on the floppy. 🤔
Soundgarden’s Badmotorfinger is underrated. Twenty five years later, still no skips.