Hello again! It’s nice to see you here!

I suspect many folks in tech have had this experience: Something that you've struggled with in the past or even put off doing because there's something tricky about it or something that is going to require a few hours of research is instantly solved or made much easier by a new tool.

For me in the past year or so this has ranged from minor whimsies and conveniences from the past year to long-delayed or denied software products and passion projects, to managing my schedule, my priorities, my business.

The whimsical ones have scratched itches...

• a simple web app that shows us up-to-the-day the annual rainfall tied to the rain season in California and not the calendar year

• a script running on one of my machines that automatically detects and blocks a recurring "sockpuppet" troll on a community site I read

• another script that prints the New York Times crossword puzzle every day, reformatting it on Sunday if necessary, and also sends a copy to my reMarkable tablet

The ambitious ones have started to take up all my time:

• a product management "apprentice" agent who will learn how you work and keep all that knowledge with you

• a group chat tool for agent conversations

• a to-do list apps that makes you focus on one thing at a time ("you had one job!")

• an animation showing tectonic plate movement over billions of years and the sequential formation of supercontinents

• trust me, there are more

... but all of these "shortcuts" share one thing in common. The moment that they solve my problem, save me time, or abstract away something I was expecting to have to do myself manually, I immediately move my own goalposts, ask myself "OK, now what?" and pile on yet another thing.

My pal Ted was interested in how I work with AI agents and we met and shared screens. The moment the agent was dispatched to work on something he suggested that it made sense to immediately use my free attention to start another agent working on something else. I agreed! This is indeed the temptation and it is absolutely the pattern I fall into over and over, amping up the hedonic treadmill.

AI made building cheap. Fast, too.

It's not like I'm pointing out some secret here. But when everybody and everything gets the same speed boost, then the race just continues on another level. Acceleration really is not the interesting part of what's going on.

When building your ideas was expensive, the cost constraint actually provided a form of discipline. It was a filter. You couldn't afford to build the wrong thing, so scarcity did your prioritizing for you (or you found out the hard way that it should have). I'm not saying it forced you to make good decisions or that politics wouldn't factor in, but it was unavoidable. Every feature had to fight for budget, and a lot of bad ideas died in that fight before anyone wrote a line of code. The expense was a tax, sure, but it also worked as a gate.

Now the floodgates are open. You don't just get to build the good things faster. You get to build everything, including all the things that used to die in triage, plus every new idea anybody has today or tomorrow. You can now waste an astonishing amount of effort, very efficiently, at scale, and feel productive the whole time because look at all this stuff getting shipped!

So what happened to the old bottleneck? Is it gone? No, it just moved. It moved upstream (OK, mixing metaphors here but stay with me), to knowing what's actually worth building, and that part has absolutely not gotten any cheaper. If anything the flood made it harder, because now there's nothing standing between you and the wrong idea except your own judgment about what's worth doing. And it's pretty hard to think with all these noisy ideas popping off all day long.

Let's just say there's a lot of software being shipped today with zero users.

And this brings me to the uncomfortable part, the part where I don't have a tidy product answer already figured out and tied in a bow for you.

Knowing what's worth building is a judgment problem. And judgment doesn't scale like compute. You can 10x your output overnight by some definitions, but you cannot 10x your taste overnight (if that's even possible), or your understanding of the people you're building for, or your willingness to look at something you spent two weeks on and say it shouldn't exist. Those things come from context, from having been wrong before and remembered why, and from caring about the outcome more than the artifact.

I've been grappling with how to use AI tools since I worked on a mental health community product and we trained a bot using NLP (basically a generation less clever than today's LLMs). Even then the venture capitalists sniffing around wanted us to replace all the people volunteering to support other real humans with AI. I never thought that was the answer. I wanted to use the AI to coach and offer suggestions to the humans offering the support, giving them superpowers, somewhat akin to a grammar checker ("are you sure you want to say it that way?").

The trick is to find the right division of labor. What are computers good at and what are we good at? That might be changing, but there are some essentials. They are better at math, repetitive processes, and so on. We are better at ... well understanding people and their needs.

I don't have much difficulty picturing how I want to work with AI tools. I want them to do all the mind-numbingly boring stuff and the things I am no good at or that would take me way too long to get the same result.

I don't want them doing the fun stuff, taking away my creative choices, making judgments or decisions about what matters, coming up with the vision, or deciding how to treat people. I know the parts I want to do and much of the operating model I've developed for myself iterating over the past 18 months or so is geared toward surfacing the parts of my work that need my attention and my discernment, and getting me out of the business of busy work and back to the hard, righteous work of deciding, describing, and refusing.

So I'm obviously not saying the use of AI is inherently bad or harmful (well, aside from the toxic effects of a broligarchy ruthlessly pushing it as an addictive panacea only they can be trusted to mind). I'm in it up to my elbows! What I am saying is that, just as in all those folktales where any use of magic comes with a price, the hyperacceleration of software development comes with a bill attached.

So what do you actually do about it? Here's a few things that I tend to recommend:

1. Before you bring AI or agentic tools into anything you're working on, figure out the core human part, the experience, the impact, how you expect to participate, where other people may or may not be involved. Think it through. Discuss it, write it down.

2. Then when you do start to introduce AI research or agents into the picture, clearly define which aspects of the work will be farmed out and which will remain in your hands or those of other people on the team.

3. Next, structure your tooling and your operating model to surface decisions, ambiguities, dilemmas, judgment calls, approvals, and so on so that bots working for you can proceed unhindered when not requiring human oversight and also appropriately pause or park work and move onto something else when faced with something "above their pay grade."

4. And finally, don't let anything leave your hands that you haven't reviewed carefully for the factors that matter to you and for which you are accountable. "The AI did it" is not an excuse.

I'm starting to develop shorthands for these habits, by the way. I have always felt that "human in the loop" sounds like something between an afterthought and a copout (plus something about HITL is just bad vibes?). So I'm using these concepts for now:

• Human at the start (HATS)

• Human owns the loop (HOTL, pronounced hotel)

• Human as last test (HALT)

Try them on for size.

Making the right decision is the work now and, well, it was always the work if we're being honest. We just can't hide behind cost now to make our tough calls for us. When the code was the hard part, you could confuse action with progress and get away with it for a while. You can't anymore.

It's so cheap to build that it becomes painfully obvious when you're just running in place, getting nowhere. The only thing left to be good at is knowing what to build and having the nerve to skip the rest.

That's the now what.

xian

—

P.S. If this is the part of your work that's gotten harder and not easier lately, that's not a you problem, it's the whole shape of the moment. Hit reply and tell me what you're wrestling with. I read all of them.

P.P.S. I will soon be making myself available for Office Hours, inviting people working with AI or agents to drop in and get help with wherever you currently are on your own learning curve. Stay tuned for details!