'Cattle, not pets' for code (ryanmadden.net)

Cloud-era “cattle, not pets” thinking—treating infrastructure as disposable and automated—may soon apply to code as LLMs dramatically lower the cost of generation. The piece argues that when code is cheap to produce, teams can stop treating every line as precious and instead optimize systems that reliably generate, verify, and replace code. That shift is significant because it reframes engineering from handcrafted artifacts to managed outputs: focus on the machinery that encodes intent, tests, and style rather than on individual implementations. The result could unlock faster iteration, cheaper maintenance, and new automated development workflows at scale. Technically, the author proposes a three‑pillar framework—durable intent (English/spec files checked into source control and versioned), automated tests as context-preserving validators, and policy/style manifests (e.g., AGENTS.md)—with LLMs orchestrating generation. Key gaps today are ephemeral chat-based intent, LLM context-window limits, and weak observability for agentized workflows. Practical remedies include using diffs of intent as LLM inputs, subagents to run tests and style checks in separate contexts, agent observability (session replay, A/B testing), and tighter config-as-code integration. Risks remain—spec precision, test deletion by models, and context scaling—but the combination of distributed specs plus LLM-powered spec-checking and multi-agent tooling could make “code-as-cattle” a plausible, highly automated engineering paradigm.