Pencils Down, Notation Up

The piece argues that large language models and agentic toolchains will inevitably push compilation targets ever lower (Rust, C, assembler), but the real decision is what to keep as the source of truth at the top of that drill. Using a concrete experiment - compiling Basecamp’s Campfire Rails app to a single 202k-line C file - the author shows why maintaining the high-level Rails code is far superior for agent-driven edits. The Rails source fits entirely in an agent’s context, expresses each intent once (e.g., has_many :boosts), preserves human-readable intent, and is something models have seen before; the compiled C is enormous, duplicates hundreds or thousands of low-level edits, loses intent information, and is impractical to edit or verify.

The author lays out five concrete requirements for a source of truth - fits in context, single expression of intent, preserves intent, deterministic reproduction, and model familiarity - and argues that precise prompts would effectively become a new DSL if used as the authoritative specification. Benchmarks show that compiling Rails to native code can yield large throughput and memory wins, so performance and keeping Rails as notation are compatible. The conclusion: let the agentic “drill bit” go deep for performance, but retain compact, conventional Rails notation at the top as the maintainable, trustworthy specification.