Stripe predicted at Sessions this year that “2026 Q1 will be looked back on as the first quarter of the singularity”. This year has truly marked the beginning of a remarkable time. As I embark on a new chapter in my career as a solo builder, I marvel at how fortunate I am to have aligned myself with a great new chapter for building internet businesses. Leaning into this chapter means lowering your threshold for trying things, and even letting yourself get a bit manic. Restrained AI psychosis is a perfectly reasonable reaction to this remarkable time we live in, when the cost of transforming ideas into reality has fallen to the floor.

With a sufficiently advanced software factory, you can print & operate software at industrial scale. Distribution of all that software is a separate challenge (still figuring that part out). My nature, as an engineer, is to begin by optimizing the means of production.

The throughput of your factory depends on two things: the rate of input and token efficiency.

To improve the rate of input to my factory, I’ve focused on 3 areas:

  1. Autonomy. I aim for my agents to be capable of working for hours or days off a single short prompt. Some iterative planning and specification is still required when I start large initiatives, but most of my steering happens at arms length.
  2. Routines. I monitor a growing herd of agent routines. These require much less input from me: they iterate reliably toward some goal, evolving themselves over time.
  3. Inspiration. I drink from a firehose of ideas so that I have plenty of high quality input for my factory. I’ve found X and HackerNews to be good watering holes, and I’m building a network of like-minded friends. I frequently feed my factory with the best open-source projects, blog posts, and research papers I encounter.

To improve token efficiency in my factory, I’ve focused on 2 areas:

  1. Agent productivity. Your developer platform plays a pivotal role in the output of your factory. It’s well understood in software companies that strategic investments in shared foundations can increase velocity; they call it “developer productivity”. Agent productivity is the new developer productivity. Many old tools we used for optimizing human engineers are still relevant, but the field of agent productivity is brand new.
  2. Agent coordination. Human teams pay a heavy coordination tax: scaling headcount doesn’t predictably increase output. Agent teams should in theory scale much more predictably, but it’s hard to get right. I have some experience here (I was #3 on the Tokscale leaderboard last week), and I believe shared memory is a core part of the solution.

Most pursuits boil down to modeling your domain, routing information, and iteratively building toward some fixed goal. If the means of production can be automated, put on your hard hat – it’s time to build a factory.

⌁ Ben