"Claude Is Not a Compiler" argues for treating coding agents as non-deterministic

Sources: blog.exe.dev, HN discussion
Summary: A practitioner post argues that treating a coding agent like a deterministic compiler, expecting the same prompt to yield the same correct output, is the wrong mental model, and that workflows should be built around variance, verification, and constrained scope instead. It drew a large Hacker News thread on 2026-07-21.
Why it matters: How teams frame agent reliability shapes whether they add verification and guardrails or ship unreviewed generated code.