Software That Survives
I didn't set out to write a Steve Yegge commentary blog. But he's a year ahead of me on this journey, he communicates well, and Beads transformed my agent workflow from repeated project collapses to repeated successes. So: more Yegge references incoming. You've been warned.
He just published Software Survival 3.0, a framework for predicting which software survives when AI can write anything on demand. The core insight is simple: in a world of constrained compute, software survives if it saves cognition.
He gives six levers for survival. The first two are about saving tokens directly: compress hard-won insights (Git embodies decades of wisdom about version control that would be insane to rediscover) and use cheaper substrates (grep runs on CPUs, not GPU inference — orders of magnitude more efficient for pattern matching). The next two are about reach: be broadly useful (tools that apply everywhere amortize their learning cost) and be findable (agents have to know you exist, whether through training data or documentation). The fifth is about usability: minimize friction by building "desire paths" — implementing whatever agents try to do until their guesses become correct. And the sixth is the wildcard: the human coefficient, where software survives specifically because humans want humans involved.
Tools with high ratios — where using them saves way more tokens than knowing and operating them costs — get what Yegge calls "plot armor." grep isn't going anywhere. Neither is Git.
Reading this, I realized I've been circling these ideas for weeks without having the language for them.
When I wrote about running tighter ships for agents than we ever ran for ourselves, I was describing Yegge's first two levers: insight compression and substrate efficiency. Strict types, linters, tests — they're crystallized cognition about what makes code correct. And they run on CPUs, not GPU inference. They catch mistakes in a cheaper substrate than letting the model rediscover them through trial and error. Every lint rule that fires is tokens saved.
When I wrote about the year of the agent swarm, I was describing tools that score high on his ratio. Beads survives because it compresses decades of insight about task tracking into something agents can query without re-inventing project management from scratch. Gas Town survives because orchestration is genuinely hard, and the alternative — building your own factory from first principles — would burn tokens like kindling.
The framework also explains the industrialization framing I borrowed from Subbu Allamaraju. Selection pressure is industrialization. Inefficient software gets routed around. The craft becomes commodity not because anyone decided it should, but because the economics select against waste.
What survives isn't necessarily what's elegant or beloved. It's what saves cognition relative to the cost of knowing and using it. That's a different filter than we're used to applying.
The question I keep coming back to: am I building things that would be crazy to re-synthesize? Or am I building things an agent will route around the moment it gets slightly smarter?
I don't have a confident answer yet. But at least now I have a framework for asking.