My Agent Stack for Automating My Personal Life
Plus how my air conditioner ended up in the New Yorker, the camera vs. the engine, and why thin stopped being a flex.
I’m writing this on a Monday morning after one of those early-summer weekends that’s hard to improve on: a round of golf Saturday morning, a long afternoon on the floor with my son, a date night with my wife, and hanging with friends at home then the pool on Sunday. My toddler just learned to say “I love you” unprompted, and it is absolutely wrecking me. Hope your summer/winter has been off to an equally good start.
For the last six months I’ve been building what I’ve started calling an AI business brain, a personal operating system on top of Claude Code that more and more of my work now runs on. It’s changed how I work more than anything has in years (probably ever?)
I just published the first of a few pieces on it, A Lever Made of Agents, that looks at why it’s been so impactful. It goes into why agentic AI is a new source of operating leverage and why the durable advantage comes from process power, not prompts. Two more are on the way in the next few weeks: a more practical build guide and a case study of how I use it.
I’ve started working with companies to implement this system. If you run a business and want help finding the highest-leverage places to put AI to work in your business, you can find out more here.
“No other area offers richer opportunities for successful innovation than the unexpected success.” — Peter Drucker
Why AI Didn’t Transform Our Lives in 2025
by Cal Newport (Archive.is Free Version)
I talked to Cal Newport for his New Yorker column about why AI didn’t upend everything in 2025. My wife thinks it hilarious that it’s built around how I used Claude Code to replace our HVAC system.
I used this example because I think it’s actually a great use case for thinking about how knowledge work broadly (rather than just software engineering) can be improved by AI.
I had a 27 year old HVAC system that needed to get replaced and a stack of HVAC quotes that read as gobbledygook to me. I pointed an agent at the PDFs and had it pull the actual unit models, find the manuals, and started work out what to do. It took a few sessions and a lot of supervision, but I came out with a better system at a better price than if I’d just picked a quote.
It’s a good AI use case because picking an HVAC system kind of looks like a lot of knowledge work. There is not one “right” answer and, unlike with code, you can’t just cheaply build a version and test it. Getting a “good solution” for an HVAC system is pretty similar to making a “good presentation” at your job or giving a customer a “good customer service experience.”
The job requires thoughtful research and good project management: read some quotes, track down some manuals, cross-reference the specs against the actual constraints of my house, and synthesize it into a logical recommendation given the context.
For a human, that’s an enormous amount of unfun stuff so most people just sign a quote from whoever gives the best vibes. Once you have an agent that never gets bored and can read a vast amount in minutes, suddenly the task is mostly just steering which is both faster and more fun.
Once I saw how much of my real work has that same shape (Taking lots of information, coming up with a logical synthesis and moving it forward), I started building the same setup deliberately, at the scale of the whole business. That’s what became the AI business brain, and it’s the thing I’m most excited to be working on right now.
A Camera, Not an Engine II
by Venkatesh Rao
Venkat has a frame for AI that I quite like. We keep talking about AI as an engine, but the more interesting work is camera-like: an instrument for seeing in latent space, not an engine of production.
An engine takes stored energy and turns it into motion. A camera shows you something you couldn’t see before.
Using AI like a camera is when you point AI at a domain you only half understand and use the feedback to learn its shape. That’s some of the most interesting and fun experience I’ve had with AI. I wrote an article years ago about how we should optimize for interesting. It’s a summary of some research arguing that the evolutionary rationale for the emotion ‘interesting’ is that we experience it when we are working on something where we have the potential to make rapid progress in learning.
Interesting is the feeling of compression progress: the moment a pile of messy detail collapses into a simpler pattern you can hold in your head and you feel like you gained real insight.
That’s exactly what a camera-like tool is good for. There’s a huge amount of latent knowledge sitting around what you’ve already read and half-thought, and these tools are unusually good at helping you surface it. The bonus with agentic AI is that you can build the thing out as you go, so each pass leaves you with a little more structure than the last. I’m learning faster than I have in a long time and that’s the part I think is really fun.
When Action Beats Prediction
by Cedric Chin
Cedric Chin has a good summary of Saras Sarasvathy’s work on effectuation, which is a fancy word for a simple entrepreneurial pattern: in highly uncertain domains, action often beats prediction. It is the best theoretical model for entrepreneurship I know of and a thread I’ve pulled on before.
The standard strategy model says you forecast the market, pick the best expected return, and execute. In practice, good entrepreneurs often do something different. They start with who they are, what they know, and whom they know. Then they take small actions, recruit committed stakeholders, and let surprises become data.
Causal reasoning asks, “What’s the expected return?” Effectual reasoning asks, “What can I afford to lose?” It rhymes with some of the ergodicity/Kelly-betting thread for sizing investments that I’ve written about. Both approaches suggest not making one giant forecast-dependent bet, but thinking about a series of non-ruinous ones that expose you to upside and teach you something about the territory.
Planning isn’t bad and forecasting isn’t useless, but the best measurement instrument is often a cheap action. E.g. Try to sell the thing before you spend too much time on the spreadsheet.
One observation I have about good entrepreneurs is that they both tend to seek out surprise AND they tend to re-orient around it much faster than most people. To paraphrase Peter Drucker: unexpected success is the best source of strategy.
After Automation
by Dan Shipper
You’ve probably seen charts of AI models rapidly scaling benchmarks accompanied by scary quotes about 50% of all jobs being eliminated. Dan makes a point that I think is central to the next phase of AI work: benchmarks happen inside frames.
If you look at the actual questions used on these benchmarks (and there are some examples in the article), you will notice that they are very specific and thoughtfully framed questions. They layout what data is relevant, how to think about, and what a good output looks like.
For most knowledge workers, that’s the hard part of work!
A benchmark can tell you whether a model is getting better at a specific task. That’s useful, but it hides that someone had to decide what problem mattered, what counted as success, where the boundary was, and what tradeoffs were acceptable. It is absolutely true that the models can handle more and more expansive frames, but there’s always some framing going on.
This is why I’m skeptical of simple “AI can do X percent of jobs” arguments. The framing is often doing more work than the model. Give Claude a well-specified audit task with clean spreadsheets, explicit constraints, and a narrow output format, and it may perform very well. But a lot of “intelligence” went into making the problem look like that in the first place.
The strongest version of the argument grants that models can help with framing. They can, and I use them for that constantly. The point is that someone still has to decide which frame matters, when it’s exhausted, and what counts as a better one. That kind of judgment is what David Chapman calls meta-rationality, and it’s where a lot of the value is moving.
Alex Imas and Phil Trammell: What Remains Scarce After AGI?
Dwarkesh Podcast
Labor share’s share of income has stayed surprisingly high (always around 60%) through centuries of automation. Machines changed farming, manufacturing, accounting, logistics, and office work, but the economy didn’t simply become “capital earns everything, labor disappears.”
The explanation I most like for this is that jobs are bundles of tasks and technology changes the bundle and re-assigns values but doesn’t make it go away.
Being a surgeon requires many different skills and doing lots of tasks. You need to have good fine motor control, you need relevant medical knowledge, and it’s nice if you have a pleasant bedside manner. It’s also true that a lot of medical work is paperwork, insurance calls, forms, notes, and care coordination.
If AI automates some of those tasks, the remaining work may become more valuable, not less. The same person might spend less time filling out paperwork and more time doing relational tasks and using their fine motor control.
Benedict Evans makes a similar point about the unpredictability of how technologic impacts jobs in his piece on AI job exposure. We’ve spent a century automating accounting with computers. Today, we have more accountants than ever. Job titles are often bad maps of the work underneath them and making one part of a job easier may also make other parts of it more in demand. I would love it if my doctors spent more time talking to me and less time filling out forms.
None of this means AI will be painless for labor. The messy middle could be very messy. But “task exposure” and “job replacement” aren’t the same thing and most forecasts blur that distinction.
My Agent Stack for Automating My Personal Life
by Nicolas Bustamante
For me, the useful takeway in this thread is that an agent can do the glue work across tools, not just answer questions.
Nicolas gives the example of a friend asking for an intro. The real work was scattered across WhatsApp, Gmail, Google search, job postings, email drafting, and a follow-up text. None of the steps were hard. They were just context-switchy and time consuming. The agent read the thread, found the contact, researched the company, drafted the intro, and waited for approval.
That’s the pattern I think most businesses should be looking for as no brainers right now and rhymes with the HVAC example I gave earlier. The best early AI workflows often aren’t glamorous. They’re the annoying cross-service and cross-context workflows where attention leaks out through app switching: email to calendar to Drive to CRM to browser portal to Slack. The work isn’t complex, but it’s expensive because it fragments the day.
The security caveat matters just as much as the workflow. Once an agent can read private messages, browse the web, and send things, you’ve built something powerful enough to hurt you. My own bias is to keep the line very clear: agents can read and draft; humans send.
The Millennial Midlife Crisis Is Going to Be a Barbell
by Priyanka Desai
Priyanka Desai’s piece is a very online, New York-coded theory of millennial midlife, but a lot of it hits for my less online and less New York social reality.
Technological abundance tends to change what’s high status. If one signal gets cheap, the market searches for a costlier one.
For a long time, thinness was a status marker because it implied discipline, genetics, money, or all three. GLP-1s change that. If appetite can be pharmacologically muted, then being thin becomes less impressive. Status moves to what’s still hard to fake: muscle, VO2 max, marathon times, ritual, membership, real-world intimacy. Protein-maxxing, run clubs, and deliberately offline rituals are all versions of this.
The weirdest example in the piece is South Korean “dopamine sites,” where people browse fake restaurant menus, fill carts they won’t check out, and simulate consumption without acquisition. I kind of like this theory that we’re all culturally downstream from South Korean teenagers. It’s both horrifying and hilarious in a Douglas Adams meets Black Mirror sort of way.
If you’re active on any of these platforms and want a daily(-ish) thought between newsletters, you can follow me on Substack Notes, Twitter/X, or LinkedIn.
I've had several discussions with Claude about the type of work it finds "interesting," why it finds that type of work interesting, and what the concept of interesting actually means to it (an entiry with no mechanism for emotion).