Sunny

← All weeks

What I’ve been reading (W31, 2026)

How agents change the software factory, what still needs human design, intent-driven interfaces, and why CPUs spend their time waiting

  1. The real software factory is not everyone using coding agents; that is the 2x. The 10x is when agents can take feedback, change the code, spin up a test environment, run evals, review the result, and safely merge low-risk work without waiting for a human at every step, while humans keep owning incidents, auth, billing, and other critical paths. The metric is not lines of code or PR count, but how few human touches it takes to safely ship a change. x.com/hamidships/status/2082656041267822828
  2. OpenAI is a place where things move very fast: prototypes matter more than design docs, Slack matters more than email, and you can go for weeks with barely a meeting. The infrastructure is not always as polished as Google’s, but people are trusted to figure things out, use Codex when it helps, ignore it when it does not, and just build the thing instead of asking permission first. mihai.page/six-months-at-openai
  3. A design doc should help you make the expensive decisions before they become expensive code. Focus on the choices that are painful to reverse, like architecture, interfaces, storage, dependencies, and security; give readers enough context to understand the problem, spell out goals and non-goals, and turn vague requirements into measurable targets. Keep open questions and rejected alternatives in the doc so the team can challenge the design before implementation starts. refactoringenglish.com/excerpts/write-an-effective-design-doc
  4. Software is built around what computers can do, not how humans naturally think. Windows, icons, menus, and pointers expose those capabilities, but force people to learn where things are and what each control means. AI reverses this: you simply describe what you want, and the software figures out how to do it. x.com/LukeW/status/2080330415005708476
  5. A lot of CPU performance problems are actually data movement problems. The CPU can execute instructions extremely fast, but it spends much of its time waiting for data to arrive from memory; Drepper’s classic argues that memory access, not compute, is the limiting factor for most programs, which is why CPU caches exist and why they still need help from the programmer. people.freebsd.org/~lstewart/articles/cpumemory.pdf