Steve (Builder.io)

My granola notes on software engineering 👇 Tactics vs. Principles in Software Engineering •Most of software engineering is tactics and heuristics, not principles ◦Pull requests, CI, testing, linters, git hygiene: all heuristics ◦Heuristics evolved when code was expensive; that assumption no longer holds •True principles are rare; simplicity is one that survives any era ◦Simplicity: fewer rules, fewer exceptions, fewer moving parts ◦Distinct from “easy”: a minimal CPU architecture is simple but not easy to build on ◦The tension between simplicity and easiness is a core engineering challenge •DRY, SOLID, clean code: heuristics built for humans reading code, not principles ◦Humans may no longer read or review most code; these rules deserve reexamination AI Has Removed the Speed Limit, Not Just Raised It •A single developer can now generate billions of lines of code per day ◦Not 10x or 100x faster: the ceiling is gone •Raw output volume is not the goal; trade-offs still govern everything ◦Speed up one knob, something else must compensate •Removing human code review requires lifting other things: ◦More testing, more observability, more linting ◦New frameworks that scope and contain code better ◦Architecture discipline to prevent “metastatic” bad code spreading across the codebase Architecture Over Code •Code is cheap and disposable; architecture is what compounds ◦Goal: be able to delete a feature folder and a million lines of code cleanly tomorrow ◦Coupling is still a genuine principle: contained code is safer than code that infects everything •Human engineers’ job is now to get the architecture right ◦Understand long-term direction: on-prem vs. cloud, security posture, multi-year feature roadmap ◦Build an architecture that lets the team move fast sustainably •Moving fast is no longer optional: it is table stakes for survival as a product company The Centaur Problem: When Human Input Becomes the Bottleneck •Human + AI is not always better than AI alone ◦Chess analogy: centaur teams (human + AI) beat both separately, until computers surpassed them entirely ◦Human input now makes chess engines worse; the same shift is coming in engineering •The Loom/Figma example: months of human-babysitting AI produced a mediocre 20% result ◦One person on Twitter used a well-structured agent harness over a weekend and replicated Photoshop-level work ◦The human babysitting the loop was the slow part •Engineers must identify where the moving target is and ensure they add value, not drag PR Culture and the New Engineering Contract •The economics of PRs are broken in an agentic world ◦Agents can shatter PR scope instantly; async review puts all burden on the reviewer •Shift to self-review: engineers own and are judged by their own PRs ◦Agent does the work; the engineer reviews and is accountable for what merges •AI code review (e.g. Reptile, Cursor) can steer quality without human meat-proxying •Open source mentoring heuristic no longer applies ◦Previously: invest time helping a contributor grow into a community member ◦Now: an agent-submitted PR either merges on merit or it does not; the growth loop is gone •The engineering challenge is building infrastructure that absorbs agentic velocity without the product falling apart